Data object push method and device

Abstract

This application discloses a method, apparatus, and device for pushing data objects. The method includes: acquiring social relationship information of a target user, interaction relationship information between the user and a data object, a first feature of the user, and a first feature of the data object; acquiring a second feature of the user and a second feature of the data object based on the interaction relationship information, the first feature of the user, and the first feature of the data object; acquiring the similarity of the second features between users based on the second feature of the user; adjusting the social relationship information based on the adjusted social relationship information; adjusting the second feature of the target user based on the adjusted social relationship information; and determining the data object to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object. This processing method introduces social relationship adjustment into social recommendation, improving the quality of user understanding by improving the user's social relationships, thereby enhancing the effectiveness of social recommendation.

Description

Technical Field

[0001] This application relates to the field of data processing technology, specifically to a data object push method and apparatus, a social recommendation model processing method and apparatus, and electronic devices. Background Technology

[0002] With the development of recommendation systems, social recommendation methods have attracted increasing attention. The core idea of ​​social recommendation is to improve recommendation effectiveness by incorporating rich social relationships between users. Traditional social recommendation relies on users' existing social relationships. However, existing social relationships are often noisy and sparse, making it difficult to achieve satisfactory results by directly utilizing them. Summary of the Invention

[0003] This application provides a data object push method to address the problem of poor social recommendation performance in existing technologies. This application also provides a data object push apparatus, a social recommendation model processing method and apparatus, and an electronic device.

[0004] This application provides a data object recommendation method, comprising: acquiring social relationship information of a target user, interaction relationship information between the user and a data object, a first feature of the user, and a first feature of the data object; acquiring a second feature of the user and a second feature of the data object based on the interaction relationship information, the first feature of the user, and the first feature of the data object; acquiring a second feature similarity between users based on the second feature of the user; adjusting the social relationship information based on the second feature similarity; adjusting the second feature of the target user based on the adjusted social relationship information; and determining a data object to recommend to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0005] Optionally, obtaining the second feature similarity between users based on the user's second feature includes: obtaining the second feature similarity between the target user and social users based on the user's second feature; adjusting the social relationship information based on the second feature similarity includes: identifying noisy users among the social users based on the second feature similarity between the target user and social users; and deleting the noisy users from the social users.

[0006] Optionally, the second feature similarity between the target user and the social user is obtained by the following steps: obtaining the social relationship feature between the target user and the social user; obtaining the third feature of the social user based on the social relationship feature and the second feature of the social user; and obtaining the second feature similarity between the target user and the social user based on the second feature of the target user and the third feature of the social user.

[0007] Optionally, obtaining the second feature similarity between users based on the user's second feature includes: obtaining the second feature similarity between the target user and highly active users based on the user's second feature; adjusting the social relationship information based on the second feature similarity includes: selecting a target highly active user from multiple highly active users as a new social user of the target user based on the second feature similarity between the target user and highly active users.

[0008] Optionally, it further includes: obtaining representative high-active users of multiple high-active user interest categories; obtaining the second feature similarity between the target user and the high-active users based on the second feature of the user includes: obtaining the second feature similarity between the target user and the representative high-active users based on the second feature of the user; selecting a target high-active user from multiple high-active users based on the second feature similarity between the target user and the high-active users includes: selecting a target representative high-active user from representative high-active users of multiple high-active user interest categories based on the second feature similarity between the target user and the representative high-active users.

[0009] Optionally, adjusting the second feature of the target user based on the adjusted social users includes: obtaining a seventh feature of the target user influenced by the target high-activity user based on the second feature similarity between the target user and the target high-activity user, and the second feature of the target high-activity user; obtaining an eighth feature of the target user based on the second feature and the seventh feature; and obtaining the adjusted second feature based on the eighth feature.

[0010] Optionally, obtaining the adjusted second feature based on the eighth feature includes: obtaining the sixth feature of the target user based on the first feature of the target user; and obtaining the adjusted second feature based on the sixth feature, the second feature, and the eighth feature of the target user.

[0011] Optionally, it further includes: obtaining representative highly active users from multiple highly active user interest categories; obtaining the second feature similarity between users based on the user's second feature includes: obtaining the second feature similarity between the target user and social users, and the second feature similarity between the target user and the representative highly active users based on the user's second feature; adjusting the social relationship information based on the second feature similarity includes: identifying noisy users among the social users based on the second feature similarity between the target user and social users, and deleting the noisy users from the social users; selecting a target representative highly active user from the representative highly active users from multiple highly active user interest categories based on the second feature similarity between the target user and the representative highly active users. The process involves: 1) adding new social users as the target user; 2) adjusting the second feature of the target user based on the adjusted social users, including: obtaining a fourth feature of the target user influenced by the remaining social users based on the second feature similarity between the target user and the remaining social users after removing noise users, and the second or third feature of the remaining social users; 3) obtaining a seventh feature of the target user influenced by the new neighbor users based on the second feature similarity between the target user and the new neighbor users, and the second feature of the new neighbor users; 4) obtaining a ninth feature of the target user based on the second feature, the fourth feature, and the seventh feature; and 5) obtaining the adjusted second feature based on the ninth feature.

[0012] Optionally, obtaining the adjusted second feature based on the ninth feature includes: obtaining the sixth feature of the target user based on the first feature of the target user; and obtaining the adjusted second feature based on the sixth feature, the second feature, and the ninth feature of the target user.

[0013] Optionally, the process of obtaining the second feature similarity between users based on the user's second feature, adjusting the social relationship information based on the second feature similarity, and adjusting the second feature of the target user based on the adjusted social users can be performed multiple times.

[0014] This application also provides a method for pushing data objects, including:

[0015] Receive a request from the client to retrieve recommendation information for a data object;

[0016] Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics;

[0017] Based on the interaction relationship information, the user's first feature, and the data object's first feature, obtain the user's second feature and the data object's second feature;

[0018] Based on the user's second feature, obtain the second feature similarity between users;

[0019] The social relationship information is adjusted based on the second feature similarity.

[0020] The second characteristic of the target user is adjusted based on the adjusted social relationship information;

[0021] Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object to be recommended to the target user is determined;

[0022] Provide data object recommendation information to the client.

[0023] This application also provides a method for pushing data objects, including:

[0024] Send a request to the server to retrieve data object recommendation information;

[0025] The system receives data object recommendation information sent back from the server. The data object recommendation information is determined by social recommendation based on the adjusted social relationship information of the target user. The adjusted social relationship information is determined based on the second feature similarity between the target user's historical interaction data objects and other users. The second feature of the user is determined based on the interaction relationship information between the user and the data object, the user's first feature, and the first feature of the data object.

[0026] Displays recommendation information for the data object.

[0027] This application also provides a method for pushing data objects, including:

[0028] Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics;

[0029] Based on the interaction relationship information, the user's first feature, and the data object's first feature, obtain the user's second feature and the data object's second feature;

[0030] Based on the second feature of the target user and the second feature of the social user, the second feature similarity between the target user and the social user is obtained;

[0031] Based on the second feature similarity, noisy users among the social users are identified;

[0032] Remove the noisy user from the social users;

[0033] Adjust the second characteristic of the target user based on the adjusted social relationship information of the target user;

[0034] Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object is determined to be recommended to the target user.

[0035] This application also provides a method for pushing data objects, including:

[0036] Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics;

[0037] Based on the interaction relationship information, the user's first feature, and the data object's first feature, obtain the user's second feature and the data object's second feature;

[0038] Obtain representative high-active users across multiple high-active user interest categories;

[0039] Based on the second feature of the target user and the second feature of the representative high-activity user, obtain the second feature similarity between the target user and the representative high-activity user;

[0040] Based on the second feature similarity between the target user and the representative high-activity user, a target representative high-activity user is selected from the representative high-activity users of the multiple high-activity user interest categories as the target user's new social users;

[0041] Adjust the second characteristic of the target user based on the adjusted social relationship information of the target user;

[0042] Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object is determined to be recommended to the target user.

[0043] This application also provides a method for pushing data objects, including:

[0044] Acquire multiple interest categories of highly active users;

[0045] Based on the historical interaction data objects of highly active users and the data objects corresponding to the interest classes of highly active users, obtain the correlation between the data objects of highly active users and the interest classes of highly active users;

[0046] Based on the relevance of the data objects, obtain the interest categories of the highly active users corresponding to the highly active users;

[0047] Based on the correlation between the high-active users corresponding to the high-active user interest class and the data object of the high-active user interest class, obtain the representative high-active users of the high-active user interest class;

[0048] Based on the historical interaction data objects of the representative highly active users and the historical interaction data objects of the target users, the similarity between the representative highly active users and the target users is obtained;

[0049] Based on the similarity, representative high-active users related to the target user are selected from representative high-active users of the multiple high-active user interest categories;

[0050] Based on the relevant representative high-activity users, determine the data objects to recommend to the target users.

[0051] This application also provides a method for pushing data objects, including:

[0052] Obtain the target user's social relationship information;

[0053] Based on the historical interaction data objects of social users and the historical interaction data objects of target users, obtain the similarity of data objects between social users and target users;

[0054] Based on the similarity of the data objects, noisy users among the social users are identified;

[0055] Based on social users other than the noise users, determine the data objects to recommend to the target user.

[0056] This application also provides a social recommendation model processing method, including:

[0057] Multiple training samples are obtained; the training samples include users' social relationship information, interaction relationship information between users and data objects, users' first features, data objects' first features, and data object recommendation information.

[0058] A network structure for constructing a social recommendation model is provided. This network structure includes a feature encoder, a social graph adjuster, a feature adjuster, and a data object selector. The feature encoder is used to obtain a second feature of the user and a second feature of the data object based on the interaction relationship information, a first feature of the user, and a first feature of the data object. The social graph adjuster is used to obtain the second feature similarity between users based on the second feature of the user and adjust the social relationship information based on the second feature similarity. The feature adjuster is used to adjust the second feature of the target user based on the adjusted social relationship information. The data object selector is used to determine the data object to recommend to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0059] The parameters of the model are learned from the multiple training samples.

[0060] Optionally, the social graph adjuster includes a module for adding highly active users, used to obtain representative highly active users of multiple highly active user interest categories; obtain the similarity between the target user and the representative highly active users based on the second feature of the target user and the second feature of the representative highly active users; and select a target representative highly active user from the representative highly active users of the multiple highly active user interest categories based on the similarity between the target user and the representative highly active users, as a new social user of the target user.

[0061] The method further includes:

[0062] For the training samples of highly active users, the second feature of the highly active users is adjusted according to the feature distribution of low-active users, and the highly active users are regarded as pseudo-low-active users.

[0063] The high-activity user interest class corresponding to the high-activity user is used as the high-activity user interest class corresponding to the pseudo-low-activity user.

[0064] The parameters of the model learned from the plurality of training samples include:

[0065] The parameters of the feature encoder are learned from a plurality of first training samples, the plurality of first training samples including training samples of highly active users;

[0066] The parameters of the social graph adjuster are learned from multiple second training samples based on the model loss function, wherein the multiple second training samples include training samples from pseudo-low-activity users.

[0067] The model loss function includes a social recommendation loss term and an imitation loss term. The imitation loss term is determined based on a first similarity and a second similarity. The first similarity includes the second feature similarity between the pseudo-low-activity user and a representative high-activity user of the high-activity user interest class corresponding to the pseudo-low-activity user. The second similarity includes the second feature similarity between the pseudo-low-activity user and a representative high-activity user of other high-activity user interest classes.

[0068] Optionally, adjusting the second characteristic of the highly active users based on the characteristic distribution of low-activity users can be done in one of the following ways:

[0069] Randomly reduce the amount of information in the second characteristic of highly active users;

[0070] Adjust the second characteristic of high-active users based on the characteristic distribution of high-active users and low-active users;

[0071] The second characteristic of highly active users is adjusted to be a weighted sum of the second characteristics of low-active users and the second characteristics of highly active users.

[0072] Optionally, adjusting the second characteristic of highly active users based on the characteristic distribution of highly active users and the characteristic distribution of inactive users includes:

[0073] Obtain the second features of multiple highly active users and multiple inactive users;

[0074] Based on the second characteristics of the multiple highly active users, obtain the first mean and the first deviation of the second characteristics of the multiple highly active users;

[0075] Based on the second characteristics of the plurality of low-activity users, obtain the second mean and the second deviation of the second characteristics of the plurality of low-activity users;

[0076] Based on the first mean, the second characteristic of the highly active users, the second mean, and the second deviation, the third deviation of the highly active users is obtained;

[0077] Based on the first mean and the third deviation, the adjusted second feature of highly active users is obtained.

[0078] This application also provides a data object pushing device, including:

[0079] The data acquisition unit is used to acquire the social relationship information of the target user, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics;

[0080] The second feature acquisition unit is used to acquire the second feature of the user and the second feature of the data object based on the interaction relationship information, the first feature of the user and the first feature of the data object;

[0081] The second feature similarity acquisition unit is used to acquire the similarity between the target user and the related user based on the second feature of the target user and the second feature of the related user;

[0082] A social relationship adjustment unit is used to adjust the social relationship information based on the similarity.

[0083] The second feature adjustment unit is used to adjust the second feature of the target user based on the adjusted social relationship information of the target user;

[0084] The data object selection unit is used to determine the data object to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0085] This application also provides an electronic device, including:

[0086] Processor; and

[0087] A memory for storing a program that implements the method according to any one of the preceding claims, wherein the device is powered on and the program runs the method via the processor.

[0088] This application also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the various methods described above.

[0089] This application also provides a computer program product including instructions that, when run on a computer, cause the computer to perform the various methods described above.

[0090] Compared with the prior art, this application has the following advantages:

[0091] The data object recommendation method provided in this application involves obtaining the target user's social relationship information, the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; obtaining the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; obtaining the second feature similarity between users based on the user's second feature; adjusting the social relationship information based on the second feature similarity; adjusting the target user's second feature based on the adjusted social relationship information; and determining the data object to recommend to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature. This approach introduces social relationship adjustment into social recommendation, improving the target user's (e.g., low-activity user's) social relationships, such as removing noisy neighbors from the target user's existing social relationships and strengthening the connection between the target user and high-activity users, thereby improving the quality of the user's social relationships and ultimately improving the quality of user understanding; therefore, it can effectively improve the social recommendation effect. Attached Figure Description

[0092] Figure 1 A flowchart illustrating an embodiment of the data object push method provided in this application;

[0093] Figure 2 This application provides a schematic diagram illustrating a scenario of an embodiment of the data object push method.

[0094] Figure 3 A schematic diagram of a social recommendation model for an embodiment of the data object push method provided in this application. Detailed Implementation

[0095] Many specific details are set forth in the following description to provide a full understanding of this application. However, this application can be implemented in many other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of this application; therefore, this application is not limited to the specific embodiments disclosed below.

[0096] This application provides a data object push method and apparatus, a social recommendation model processing method and apparatus, and an electronic device. The various solutions are described in detail below in each embodiment.

[0097] With the development of social media, the fundamental reason for the effectiveness of social recommendations is that people tend to establish social relationships based on certain shared characteristics and mediums, and the formation of social networks inherently possesses a kind of network homogeneity. Existing social recommendation schemes operate on a static social relationship basis, and their upper limit depends on the given social relationships. The inventors discovered that "low-activity users" often account for a high proportion of all users, and these low-activity users not only have less individual activity but also typically have fewer friends, indicating significant room for efficiency improvement. Based on the above observations and analysis, the inventors propose introducing social relationship adjustments into the existing social recommendation system. This can effectively improve the overall performance of existing social recommendation schemes by increasing the number of users' social relationships (e.g., increasing high-quality relationships) and improving the quality of social relationships (e.g., removing noisy relationships and increasing connections with high-activity users).

[0098] First Embodiment

[0099] Please refer to Figure 1 This is a flowchart of the data object push method of this application. In this embodiment, the method may include the following steps:

[0100] Step S101: Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature.

[0101] The data object push method provided in this application is used to recommend data objects to target users. Data objects can be product objects, news objects, video objects, advertising objects, etc. Target users can be users of a data object service system (such as an e-commerce platform, video platform, etc.), specifically, they can be low-activity users or high-activity users of the data object service system. Low-activity users are those who use the data object service system less frequently, while high-activity users are those who use the data object service system more frequently. For example, users who access the data object service system less than m times per month or whose access frequency ranks in the bottom n% are considered low-activity users, while users who access the data object service system more than p times per month or whose access frequency ranks in the top q% are considered high-activity users.

[0102] like Figure 2 As shown, the data object push method provided in this application can be used in e-commerce platforms. E-commerce platforms improve recommendation effectiveness by introducing social relationships between users and leveraging the product interaction behavior of social users to recommend products to target users. Existing social recommendation schemes operate on static social relationships, and the upper limit of recommendation effectiveness depends on the given existing social relationships. On the one hand, there may be low-quality social relationships within the existing social relationships. Since users associated with low-quality social relationships have low purchase homogeneity, this affects the recommendation effect. On the other hand, due to… Figure 2It is evident that low-activity users on e-commerce platforms have significantly fewer friends than high-activity users. The number of friends determines the upper limit of recommendation effectiveness, thus the recommendation effect for low-activity users is poor based on existing social relationships. To improve the social recommendation effect for low-activity users, the method provided in this application introduces adjustments to social relationship processing into social recommendation. By improving users' social relationships, the quality of these relationships is enhanced, thereby improving the quality of user understanding and ultimately improving the social recommendation effect. For example, low-quality social relationships are deleted from existing social relationships (noisy neighbors) to improve the quality of existing social relationships; the connection between target users and high-activity users is strengthened (high-activity neighbors are added) to increase the quantity of users' social relationships while ensuring the quality of newly added social relationships.

[0103] Social relationship information includes identifying information about social users who have social relationships with the target user. This identification information could be a user's registered account on an internet platform, their mobile device ID, etc. For example, if both the target user and the social user are registered users on an e-commerce platform, the user identifier could be the user's registered account on the e-commerce platform. The social relationship between the target user and the social user could be a neighborly relationship, a colleague relationship with similar job functions, etc.

[0104] The interaction information between users and data objects can include two aspects: the set of data objects that interact with the user, and the set of users that interact with the data objects. This interaction information can be obtained based on the user's historical interaction behavior data with the data objects. Taking a product object as an example, the user's historical interaction behavior with the data object can include browsing, purchasing, favorited, and adding to cart. In practice, based on the user's historical interaction behavior data, the product objects that generated interaction behavior for each user can be aggregated, and the users that generated interaction behavior for each product object can be aggregated. For example, user "Zhang San" has interacted with 100 products including products 1, 3, 6, and 10, and users who have interacted with product 1 include 1000 users such as Zhang San and Li Si. Table 1 below shows the user's historical interaction behavior data with product objects, Table 2 shows the set of product objects that interacted with the user based on the above historical interaction behavior data, and Table 3 shows the set of users that interacted with the product based on the above historical interaction behavior data.

[0105] field name Example data Log ID number User ID Zhang San Product Object ID Sku1 Behavioral types Browse / Purchase / Favorite / Add to Cart, etc. Operation time Timestamp …

[0106] Table 1. Historical Interaction Data of Users with Product Objects

[0107] field name Example data User ID Zhang San Product object set Sku1, Sku3, Sku6, Sku10,… …

[0108] Table 2. Set of Product Objects that Interact with Users

[0109] field name Example data Product Object ID Sku1 User set Zhang San, Li Si, Wang Wu, ... …

[0110] Table 3. Set of users who interacted with the product

[0111] The primary characteristic of a user is an interpretable, explicit characteristic. This can be user attribute characteristics, user behavior characteristics, or a combination of both. User attribute characteristics can be static features, determined based on basic user information, such as region or age group. User behavior characteristics can be dynamic features, determined based on user behavior information, such as statistical data on browsing, favorites, add-to-cart, and transactions over the past month.

[0112] The primary characteristic of a data object is an interpretable, explicit characteristic. A data object can be defined using a set of attributes. This primary characteristic can be an attribute feature of the data object, which can be a static characteristic determined based on the basic information of the data object. Taking a product object as an example, its attributes include product title, product description, price, place of origin, and category. In practice, the primary characteristic of a data object can also be a multimodal description, such as images or videos.

[0113] Step S103: Based on the interaction relationship information, the user's first feature, and the data object's first feature, obtain the user's second feature and the data object's second feature.

[0114] The second feature of a user, obtained based on the interaction relationship between the user and the data object, is related to the first feature of the data object that the user has interacted with. The second feature of the data object is related to the first feature of the user who has accessed the data object. This makes the second feature of the user and the second feature of the data object contain collaborative information, and the second feature of the user and the second feature of the data object are comparable.

[0115] In one example, step S103 may include the following sub-steps: 1) obtaining the second feature of the user based on the set of data objects that interact with the user and the first feature of the data objects; 2) obtaining the second feature of the data objects based on the set of users that interact with the data objects and the first feature of the users.

[0116] In one example, step S103 may include the following sub-steps: 1) obtaining the user's second feature based on the user's first feature, the set of data objects interacting with the user, and the first feature of the data objects; 2) obtaining the data object's second feature based on the data object's first feature, the set of users interacting with the data object, and the user's first feature. This processing method ensures that the user's second feature and the data object's second feature not only contain collaborative information, but also that the user's second feature is related to the user's own characteristics, and the data object's second feature is related to the data object's own characteristics; therefore, it can effectively improve the comprehensiveness of the user's second feature and the data object's second feature.

[0117] In specific implementation, step S103 can be implemented using relatively mature graph neural networks, such as lightGCN (lightgraph convolution network) or NGCF (Neural Graph Collaborative Filtering), which will not be elaborated here.

[0118] Step S105: Obtain the second feature similarity between users based on the user's second feature.

[0119] Step S107: Adjust the social relationship information based on the second feature similarity to improve the quality of social relationships and / or data.

[0120] A user's second characteristic can be obtained based on the data objects the user has interacted with and the first characteristics of those data objects. However, low-activity users have very limited data object interactions, making it difficult for their second characteristic to reflect their potential interests. Therefore, it is necessary to supplement this with additional user social relationships to improve recommendations for target users. However, due to the noise (low quality) and sparsity (small number of social relationships) in users' existing social relationships, directly using existing social relationship information for data object recommendations cannot achieve good results. Therefore, the method provided in this application introduces the second characteristic similarity between users and optimizes and adjusts the target user's existing social relationship information based on this similarity.

[0121] Since a user's second feature is related to the data objects the user has interacted with, as well as the first feature of the data objects, and can also be related to the user's first feature, the second feature similarity can reflect the degree of similarity between users in multiple aspects, such as the degree of similarity in their preferences for data objects, the degree of similarity between the users themselves, and the degree of similarity in user behavior. The higher the second feature similarity between different users, the more likely they are to be interested in the same or similar data objects.

[0122] In one example, step S105 can be implemented as follows: obtaining the features of the user's second feature in at least one aspect; obtaining the feature similarity between the users in the at least one aspect based on the features in the at least one aspect; and obtaining the second feature similarity based on the feature similarity in the at least one aspect. The features in the at least one aspect can be interpretable explicit features, such as data object preference features, user region features, user behavior features, etc. The features in the at least one aspect can also be uninterpretable implicit features, such as projecting the user's second feature onto one or more projection heads in a neural network to obtain features processed by one or more projection heads. Using this method, a comprehensive similarity between users can be obtained based on the feature similarity between users in multiple aspects, thus improving the accuracy of the second feature similarity.

[0123] In one example, step S105 may include the following sub-step: obtaining the second feature similarity between the target user and the social users based on the user's second feature. Correspondingly, step S107 can be implemented as follows: determining noisy users among the social users based on the second feature similarity between the target user and the social users; deleting the noisy users from the social users. The target user's social users can be friends, neighbors, colleagues, etc. The target user's social users can be highly active users or inactive users. Using this processing method, noisy users whose data object preferences differ significantly from the target user's can be deleted from the target user's social users, thus effectively improving the quality of social relationships.

[0124] In one example, the second feature similarity between the target user and the social user can be obtained using the following steps:

[0125] Step S201: Obtain the social relationship characteristics between the target user and the social user.

[0126] The social relationship features can reflect the degree of social relationship between two users, such as family and friends having a higher degree of social relationship than neighbors. Step S203: Based on the social relationship features and the second feature of the social user, obtain the third feature of the social user.

[0127] To assess the value of the existing social relationship between target user i and social user j, the social relationship characteristics f between the target user and the social user can be used. ij and the second characteristic of social users Combining these, we obtain the third feature of social user j.

[0128] In practice, it can be based on the second characteristic of social user j. Using machine learning models such as multilayer perceptrons, the social relationship features f between target users and social users are analyzed. ij and the second characteristic of social users Combining these, we obtain the third feature of social user j. The third characteristic of social user j The following formula can be used for calculation:

[0129]

[0130] Where σ is the ReLU activation function, w d and b d These are the weight matrix and bias vector. || represents the feature concatenation operation, which concatenates the second features of social users. Social relationship characteristics between social users and target users f ij .

[0131] Step S205: Obtain the second feature similarity between the target user and the social user based on the second feature of the target user and the third feature of the social user.

[0132] In one example, step S205 can be implemented as follows: the similarity between the target user's second feature and the social user's third feature in a certain aspect is taken as the second feature similarity between the two users. This similarity in a certain aspect can be an explicit, interpretable similarity or an implicit, non-interpretable similarity in the neural network. Interpretable similarity can be preference similarity in terms of data object category, data object price, or data object origin, etc. Specifically, the cosine similarity between the target user's second feature and the social user's third feature in a certain aspect can be taken as the second feature similarity, as shown in the following formula:

[0133]

[0134] Among them, w n It's the projection head; the projection head determines in what aspects two users are similar.

[0135] In another example, step S205 can be implemented as follows: Based on the second feature of the target user and the third feature of the social user, obtain the similarity between the target user and the social user in multiple aspects; based on the similarity in multiple aspects, obtain the comprehensive similarity between the target user and the social user. The similarity in multiple aspects may include preference similarity in data object category, preference similarity in data object price, preference similarity in data object origin, etc. The similarity in multiple aspects may include implicit and uninterpretable similarities in multiple aspects. This processing method takes into account the possibility of multiple similarities between users, thus effectively improving the accuracy for noisy users.

[0136] In practical implementation, the above formula can be extended to a multi-head version, with each head depicting one aspect, as shown in the following formula:

[0137]

[0138] As can be seen from the above formula, the average similarity of the Q projection dimensions of the second feature is used to obtain the comprehensive similarity between the two users.

[0139] The method provided in this application embodiment, through the above steps S201 to S205, combines the social relationship features between the target user and the social user to calculate the second feature similarity between the target user and the social user. Therefore, it can effectively improve the accuracy of the second feature similarity between the target user and the social user.

[0140] In another example, the second feature similarity between the target user and the social user can be obtained by the following steps: obtaining the similarity between the target user and the social user in at least one aspect based on the second feature of the target user and the second feature of the social user; obtaining the second feature similarity between the target user and the social user based on the similarity in the at least one aspect.

[0141] In one example, determining noisy users among the social users based on a second feature similarity between the target user and the social users can be achieved as follows: Noisy users among the social users are determined based on the second feature similarity and a similarity threshold. For example, users whose second feature similarity is less than or equal to the similarity threshold are considered noisy users.

[0142] In another example, the noise users among the social users can be determined based on the second feature similarity between the target user and the social users. This can be achieved by determining the noise users among the social users based on the social user deletion ratio and the second feature similarity between the target user and the social users.

[0143] In specific implementation, determining the "noisy users" among the social users based on the social user deletion ratio and the second feature similarity between the target user and the social users can be achieved as follows: obtain the second feature similarity ranking information between the target user and multiple social users; determine the "noisy users" among the social users based on the ranking information and the deletion ratio. For example, if the deletion ratio is 30%, the social users ranked in the bottom 30% based on the second feature similarity are considered "noisy users," and the remaining social users are retained.

[0144] The social user deletion ratio can be set according to needs, such as a fixed ratio. In one example, the social user deletion ratio can be determined as follows: based on the number of data objects the target user has interacted with (the number of historical interaction data objects). Generally, the fewer data objects the target user has interacted with, the smaller the deletion ratio, and the more existing social users need to be retained. Based on this, the deletion ratio of target user u can be calculated using the following formula:

[0145]

[0146] Here, r1 is the temperature system, controlling the overall smoothness. As shown in the above formula, if the number of data objects |I(u)| that user u has interacted with approaches 0, then all of user u's social users will be retained.

[0147] In another example, step S105 may include the following sub-step: obtaining the second feature similarity between the target user and highly active users based on the user's second feature; correspondingly, step S107 may be implemented as follows: selecting a target highly active user from multiple highly active users as the target user's new social user based on the second feature similarity between the target user and highly active users. This approach increases the connection between the target user and highly active users, further enhancing the homogeneity of user purchasing behavior. Therefore, it can improve both the quantity and quality of user social relationships.

[0148] In one example, the method may further include the following step: obtaining representative high-active users of multiple high-active user interest categories. Correspondingly, the step of obtaining the second feature similarity between the target user and the high-active users based on the user's second feature can be implemented as follows: obtaining the second feature similarity between the target user and the representative high-active users based on the user's second feature. Correspondingly, the step of selecting a target high-active user from the multiple high-active users based on the second feature similarity between the target user and the high-active users can be implemented as follows: selecting a target representative high-active user from the representative high-active users of the multiple high-active user interest categories based on the second feature similarity between the target user and the representative high-active users.

[0149] A high-activity user interest class is a cluster of high-activity users based on their data object preferences. A high-activity user interest class can include multiple high-activity users and can correspond to multiple data objects. The data objects of interest to the high-activity users included in a high-activity user interest class are quite similar to the data objects corresponding to that high-activity user interest class; however, the data objects of interest to high-activity users in different high-activity user interest classes can vary significantly. Taking product categories as an example, science and technology books, casual wear, and travel packages can correspond to the same high-activity user interest category A. Users interested in these data categories are typically male researchers, so high-activity user interest category A can include highly active male researchers. Academic books and home furnishing products can correspond to the same high-activity user interest category B. Users interested in these data categories are typically mothers aged 30-50, so high-activity user interest category B can include highly active mothers aged 30-50. High-end products related to outdoor sports, such as outdoor backpacks, hiking shoes, and tents, can correspond to the same high-activity user interest category C. Users interested in these data categories are typically people with strong economic capabilities, so high-activity user interest category C can include highly active people with strong economic capabilities. Mid-to-low-end products related to outdoor sports can correspond to the same high-activity user interest category D. Users interested in these data categories are typically people with moderate economic capabilities and students who enjoy outdoor sports, so high-activity user interest category D can include highly active people with moderate economic capabilities and students who enjoy outdoor sports.

[0150] Multiple data objects corresponding to a single highly active user's interest class can span different data object categories. In other words, multiple data objects corresponding to a single highly active user's interest class can originate from different data object categories. However, multiple data objects corresponding to a single highly active user's interest class can also originate from the same data object category. Taking a product object as an example, the product object category could be books, clothing, home appliances, food, etc.

[0151] The representative highly active users are determined based on the similarity between the historical interaction data objects of highly active users and the data objects corresponding to the interest categories of highly active users. By obtaining the second feature similarity between the target user and the representative highly active users, and selecting the target representative highly active user from representative highly active users of multiple interest categories based on this similarity, the computational workload of the second feature similarity between the target user and highly active users can be effectively reduced, thereby improving the efficiency of social relationship optimization. Simultaneously, this processing method allows multiple highly active users associated with the target user to come from different interest categories. This enables the recommendation of more diverse data objects to the target user by leveraging representative highly active users from multiple interest categories, thus effectively improving the quality of social recommendations.

[0152] In one example, the method may further include the following steps: clustering the historical interaction data objects of highly active users to form the plurality of highly active user interest classes. Since user interests often reside in the data objects they have interacted with, accurately determining user interests should begin with the data objects. Highly active users are clustered according to their data object preferences to obtain various highly active user interest classes. Specifically, data object features can be clustered (e.g., K-means clustering or other clustering methods) to obtain a set C = {c1, c2, ..., c1} of highly active user interest classes, representing that the set includes l highly active user interest classes, where... This represents the set of data objects corresponding to the i-th type of highly active user interest class.

[0153] In one example, obtaining representative high-active users from multiple high-active user interest categories may include the following steps:

[0154] Step S301: Based on the historical interaction data objects of the highly active users and the data objects corresponding to the interest classes of the highly active users, obtain the data object correlation between the highly active users and the interest classes of the highly active users.

[0155] Taking product objects as an example, for each highly active user u + It can calculate the set of goods objects I(u) that it has purchased. + ) and the i-th type of highly active user interest c i The similarity between corresponding product sets (e.g., the Jaccard coefficient) is denoted as J(I(u + ), c i ).

[0156] Step S303: Based on the relevance of the data objects, obtain the interest class of the highly active user corresponding to the highly active user.

[0157] Based on the correlation between data objects of highly active users and their interest categories, highly active users u+ It is associated with the interest classes of highly active users that it matches. In practice, this can be represented by the following formula:

[0158]

[0159] Wherein, C(u) + ) represents user u + This belongs to the high-activity user interest category. Thus, each high-activity user interest category c... i They will all be assigned to some highly active users belonging to this category. Multiple highly active users corresponding to the interest category of highly active users have similar consumption habits. Let U be the set of multiple highly active users with similar consumption habits corresponding to the interest category of highly active users. + (c i )={u + |C(u + ) = c i}, U + (c i ) is the set of highly active users whose interest class is the i-th highly active user.

[0160] Step S305: Based on the data object correlation between the high-active users associated with the high-active user interest class and the high-active user interest class, obtain representative high-active users of the high-active user interest class.

[0161] The set U of highly active users in the i-th highly active user interest class + (c i Select representative users from the list. Its historical interaction data object set The set of product objects c corresponding to the i-th highly active user interest class i The similarity between them can be the largest or the second largest, etc. In practice, the representative users of each highly active user's interest category can be uniformly represented as a set. and set U C Highly active users are considered as representative highly active users, also known as anchor users, and only the possibility of establishing social relationships between target users and these anchor users is considered.

[0162] In practice, the similarity between the target user and each representative highly active user can be calculated, so that the newly added social relationships with highly active users based on their interests can enhance the feature quality of low-active users.

[0163] In one example, the second feature similarity between the target user and the representative highly active user can be achieved as follows: the similarity between the second feature of the target user and the second feature of the representative highly active user in a certain aspect is taken as the second feature similarity between the two users. Specifically, the cosine similarity between the second feature of the target user and the second feature of the representative highly active user in a certain aspect can be taken as the second feature similarity, as shown in the following formula:

[0164]

[0165] Among them, w n It's the projection head; the projection head determines the similarity between two users. (Representative high-activity users)

[0166] In another example, the second feature similarity between the target user and the representative highly active user can be achieved as follows: based on the second features of the target user and the representative highly active user, obtain the similarity between the target user and the representative highly active user in multiple aspects; based on the similarity in these multiple aspects, obtain the comprehensive similarity between the target user and the representative highly active user. This approach takes into account the potential for multiple similarities between users, thus effectively improving the accuracy of identifying highly active users with whom social relationships can be increased.

[0167] In practical implementation, the above formula can be expanded into a multi-head version, with each head representing a specific aspect, such as the user UI and a representative highly active user. Similarity between:

[0168]

[0169] in, This represents the similarity between user i and user k in aspect q. When deleting a user, as described above It is shared. It is a matrix specifically used to map anchor users.

[0170] In one example, step S107 can be implemented as follows: based on the increase ratio of highly active users and the second feature similarity between the target user and the highly active users, select a target highly active user from multiple highly active users as the target user's new social users.

[0171] In practical implementation, for the target user u i It can be based on the second feature similarity between the target user and representative highly active users. The size of the target high-active user set is determined by selecting a representative subset of highly active users.

[0172] The percentage increase in highly active users can be set according to needs, such as a fixed percentage. In one example, the percentage increase in highly active users can be determined as follows: based on the number of data objects the target user has interacted with (the number of historical interaction data objects). If a user has few purchasing behaviors, more highly active users are added. Similar to the social user deletion ratio, it is believed that low-active users need more high-quality information from highly active users, therefore, the percentage increase in highly active users for user u can be calculated using the following formula:

[0173]

[0174] Here, r2 remains a hyperparameter similar to r1. As shown in the equation above, if the number of data objects |I(u)| that user u has interacted with approaches 0, then the proportion of highly active users for user u increases by 50%. For example, for user u... i ,according to The size selection uses the top 50% of anchor users as the set.

[0175] So far, for target user u i In other words, the social user base has been improved. And the target set of highly active users, such as the target representative set of highly active users. The following will integrate them into the target user u i In the representation of social relationships.

[0176] Step S109: Adjust the second feature of the target user based on the adjusted social relationship information of the target user.

[0177] This step involves performing representation calculations on the target user based on the adjusted social relationship information of the target user, thereby obtaining the adjusted second feature of the target user.

[0178] In one example, step S109 may include the following sub-steps: 1) obtaining a fourth feature of the target user affected by the remaining social users based on the second feature similarity between the target user and the remaining social users after removing noisy users, and the second or third feature of the remaining social users; 2) obtaining a fifth feature of the target user based on the second feature and the fourth feature; 3) obtaining an adjusted second feature based on the fifth feature.

[0179] If a fourth feature is obtained based on the second feature similarity between the target user and the remaining social users after removing noisy users, and the third feature of the remaining social users, then the fourth feature can be represented by the following formula:

[0180]

[0181] in, Indicates target user u i With remaining social users u j The second feature similarity between them is used; the remaining parameters are explained in the above-mentioned formulas and will not be repeated here.

[0182] The fifth characteristic of the target user can be represented by the following formula:

[0183]

[0184] The parameters are explained in the above-mentioned formula descriptions and will not be repeated here.

[0185] In one example, obtaining the adjusted second feature based on the fifth feature can be achieved as follows: Based on the first feature of the target user, obtain the sixth feature of the target user; based on the sixth feature, the second feature, and the fifth feature of the target user, obtain the adjusted second feature. This approach ensures that the adjusted second feature includes not only the user's own characteristics, but also the user's own preferences for data objects, as well as the preferences of neighboring users for data objects; therefore, it can effectively improve the quality of the target user's second feature.

[0186] Let the first characteristic of user u be denoted as x. u The sixth feature of user u is denoted as h. u The user's sixth feature is an implicit user feature. In this embodiment, the user's first feature (the user's original feature) is projected into a common hidden vector space to obtain the user's sixth feature.

[0187] In practice, for user u, a multilayer perceptron can be used to determine the user's first feature x. u Perform the following exchange:

[0188]

[0189] Where hu is the projected representation of user u, i.e., the user's sixth feature; σ is the PReLU nonlinear activation. These are learnable parameters.

[0190] In practical implementation, skip-connection can be used to aggregate h.u e u and z u The adjusted second feature of the target user can be obtained using the following formula:

[0191]

[0192] Among them, w f and b f It is a learnable parameter; h u Indicates the sixth feature, e u zu represents the second feature of the target user (a representation encoded in the user-data object interaction graph), and zu represents the fifth feature (a representation encoded in the user's social graph). This refers to the second feature obtained after adjusting the target user's social relationships.

[0193] In one example, step S109 may include the following sub-steps: 1) obtaining a seventh feature of the target user being affected by the target high-activity user based on the second feature similarity between the target user and the target high-activity user, and the second feature of the target high-activity user; 2) obtaining an eighth feature of the target user based on the second feature and the seventh feature; 3) obtaining an adjusted second feature based on the eighth feature.

[0194] If the target high-activity users are representative high-activity users, then the seventh feature can be represented by the following formula:

[0195]

[0196] The parameters are explained in the above-mentioned formula descriptions and will not be repeated here.

[0197] The eighth characteristic of the target user can be represented by the following formula:

[0198]

[0199] The parameters are explained in the above-mentioned formula descriptions and will not be repeated here.

[0200] In one example, obtaining the adjusted second feature based on the eighth feature can be achieved as follows: Based on the first feature of the target user, obtain the sixth feature of the target user; based on the sixth feature, the second feature, and the eighth feature of the target user, obtain the adjusted second feature. This approach ensures that the adjusted second feature includes not only the user's own characteristics, but also the user's preferences for data objects, as well as the preferences of the target high-activity user for data objects; therefore, it can effectively improve the quality of the target user's second feature.

[0201] In practice, skip-connection can be used to aggregate hu, eu, and zu to obtain the adjusted second feature of the target user, which can be calculated using the following formula:

[0202]

[0203] Among them, w f and b f It is a learnable parameter; h u Indicates the sixth feature, e u The second feature representing the target user (a representation of the user-data object interaction graph after encoding), z u This represents the eighth feature (the representation of the user's social graph after encoding). This refers to the second feature obtained after adjusting the target user's social relationships.

[0204] In one example, step S105 may be implemented as follows: based on the user's second feature, obtain the second feature similarity between the target user and existing social users, and the second feature similarity between the target user and the representative highly active users. Correspondingly, step S107 may include the following sub-steps: based on the second feature similarity between the target user and existing social users, identify noisy users among the existing social users and delete the noisy users from the existing social users; based on the second feature similarity between the target user and the representative highly active users, select a target representative highly active user from representative highly active users of multiple highly active user interest categories as a new social user for the target user. Accordingly, step S109 may include the following sub-steps: obtaining a fourth feature of the target user influenced by the remaining original social users based on the second feature similarity between the target user and the remaining original social users after deleting noise users, and the second or third feature of the remaining original social users; obtaining a seventh feature of the target user influenced by the new neighbor users based on the second feature similarity between the target user and the new neighbor users, and the second feature of the new neighbor users; obtaining a ninth feature of the target user based on the second feature, the fourth feature, and the seventh feature; and obtaining an adjusted second feature based on the ninth feature.

[0205] The ninth characteristic of the target user can be represented by the following formula:

[0206]

[0207] The parameters are explained in the above-mentioned formula descriptions and will not be repeated here.

[0208] In one example, the adjusted second feature can be obtained based on the ninth feature by using the ninth feature as the adjusted second feature.

[0209] In one example, obtaining the adjusted second feature based on the ninth feature may include the following sub-steps: 1) obtaining the sixth feature of the target user based on the first feature of the target user; 2) obtaining the adjusted second feature based on the sixth feature, the second feature, and the ninth feature of the target user. This approach ensures that the adjusted second feature includes not only the user's own characteristics but also the user's own preferences for data objects, as well as the preferences of neighboring users and target high-activity users for data objects; therefore, it can effectively improve the quality of the target user's second feature.

[0210] In practical implementation, skip-connection can be used to aggregate h. u e u and z u The adjusted second feature of the user can be obtained by calculating the following formula:

[0211]

[0212] Among them, w f and b f It is a learnable parameter; h u Represents the sixth feature (the user's original feature), e u This represents the second feature (the representation of the user-data object interaction graph after encoding), z u This represents the ninth feature (a representation of the user's social graph after encoding). This refers to the second feature obtained after adjusting the target user's social relationships.

[0213] This concludes the explanation of the process of adjusting a user's social relationship information and then adjusting the user's second characteristic based on the adjusted social relationship.

[0214] In one example, steps S105, S107, and S109 are executed multiple times to obtain a higher-quality user representation. Specifically, after adjusting the second feature of the target user in the first round, the second feature e of the target user can be... u Replace with features obtained after adjusting the target user's social relationships. The process of removing noisy users and adding highly active users from existing social users is repeated to obtain a better social structure. After T iterations, based on... Obtain the final adjusted second feature E of the target user. u Based on this user characteristic, data object recommendations are made.

[0215] Step S111: Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, determine the data object to be recommended to the target user.

[0216] As described in step S103, the user's second feature and the data object's second feature contain collaborative information, and the user's second feature and the data object's second feature are comparable. Therefore, the data object recommended to the target user can be determined based on the similarity between the target user's adjusted second feature and the data object's second feature.

[0217] In one example, a social recommendation model is learned from a training sample set using machine learning. This model is then used to perform the processing steps S103 to S111 described above. The input data for the social recommendation model may include the interaction graph between users and data objects, the user's social graph, and the user's first feature x. u and the first feature x of the data object i The model's output data consists of data objects recommended to users. Social recommendation models are graph-based recommendation systems that utilize graph neural networks' message passing mechanisms in encoding the interaction between users and data objects, as well as in the subsequent integration of social information.

[0218] like Figure 3 As shown in the embodiments of this application, the method not only utilizes social relationships but also proposes a model for learning social structures for users. It introduces Graph Structure Learning (GSL) into social recommendation, using GSL technology to modify social relationships and improve their quality, thereby enhancing recommendations for low-activity users. This graph structure learning model for user social recommendation includes the following processing modules: a feature encoder, a social graph adjuster, and a feature adjuster. The feature encoder in the social recommendation model obtains the second features of the user and the second features of the data object based on the interaction relationship information, the user's first feature, and the data object's first feature. The social graph adjuster in the social recommendation model obtains the second feature similarity between users based on the user's second feature. Based on the second feature similarity, the social relationship information is adjusted. The feature adjuster in the social recommendation model, based on the adjusted social relationship information, such as deleting inefficient connections (noisy neighbor users) in user u's existing social relationships and establishing beneficial connections between user u and high-activity interest clusters, adjusts the second features of the target user. In addition, the model may also include a data object selector, which determines the data objects to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0219] like Figure 3 As shown, in this embodiment, the social graph adjuster includes a module for removing noisy users and a module for adding highly active users. The second feature of the users after one round of processing is adjusted. The second feature adjusted by the user after T-round processing. The module for removing noisy users is used to remove inefficient connections in existing social relationships, while the module for adding highly active users is used to establish beneficial connections between users and highly active users based on their interests. Figure 3 When updating the social graph structure, the model shown only focuses on removing existing social user connections and establishing connections between users with high-activity user interest groups. This avoids the overhead of calculating the probability of establishing a connection between any two users, effectively reducing computational complexity and saving computational resources.

[0220] In practice, the social recommendation model may also include a simulation learning module, which allows highly active users to simulate the distribution of low-active users, thereby improving the model's ability to model the social structure of low-active users.

[0221] As can be seen from the above embodiments, the data object push method provided in this application obtains the target user's social relationship information, the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; obtains the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; obtains the second feature similarity between users based on the user's second feature; adjusts the social relationship information based on the second feature similarity; adjusts the target user's second feature based on the adjusted social relationship information; and determines the data object to be recommended to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature. This processing method introduces social relationship adjustment into social recommendation, improving the target user's (e.g., low-activity user's) social relationships, such as deleting noisy neighbors from the target user's original social relationships and strengthening the connection between the target user and high-activity users, thereby improving the quality of the user's social relationships and thus improving the quality of user understanding; therefore, it can effectively improve the social recommendation effect.

[0222] Second Embodiment

[0223] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0224] This application also provides a data object push device, including: a data acquisition unit, a second feature acquisition unit, a second feature similarity acquisition unit, a social relationship adjustment unit, a second feature adjustment unit, and a data object selection unit.

[0225] A data acquisition unit is used to acquire social relationship information of a target user, interaction relationship information between the user and a data object, a first feature of the user, and a first feature of the data object; a second feature acquisition unit is used to acquire a second feature of the user and a second feature of the data object based on the interaction relationship information, the first feature of the user, and the first feature of the data object; a second feature similarity acquisition unit is used to acquire the second feature similarity between users based on the second feature of the user; a social relationship adjustment unit is used to adjust the social relationship information based on the second feature similarity; a second feature adjustment unit is used to adjust the second feature of the target user based on the adjusted social relationship information; and a data object selection unit is used to determine the data object to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0226] In one example, the second feature similarity acquisition unit is specifically used to acquire the features of the user's second feature in at least one aspect; acquire the feature similarity between the users in the at least one aspect based on the features in the at least one aspect; and acquire the second feature similarity based on the feature similarity in the at least one aspect.

[0227] In one example, the second feature similarity acquisition unit is specifically used to acquire the second feature similarity between the target user and the social user based on the second feature of the user; the social relationship adjustment unit is specifically used to determine the noisy user among the social users based on the second feature similarity between the target user and the social user; and delete the noisy user from the social users.

[0228] In one example, the second feature similarity acquisition unit is specifically used to acquire the social relationship features between the target user and the social user; acquire the third feature of the social user based on the social relationship features and the second feature of the social user; and acquire the second feature similarity between the target user and the social user based on the second feature of the target user and the third feature of the social user.

[0229] In one example, the social relationship adjustment unit is specifically used to determine the social user deletion ratio based on the number of historical interaction data objects of the target user; and to determine the noisy users among the social users based on the deletion ratio and the second feature similarity.

[0230] In one example, the second feature adjustment unit is specifically used to obtain a fourth feature of the target user affected by the remaining social users based on the second feature similarity between the target user and the remaining social users after removing noisy users, and the second or third feature of the remaining social users; to obtain a fifth feature of the target user based on the second feature and the fourth feature; and to obtain the adjusted second feature based on the fifth feature.

[0231] In one example, the second feature adjustment unit is specifically used to obtain the sixth feature of the target user based on the first feature of the target user; and to obtain the adjusted second feature based on the sixth feature, the second feature, and the fifth feature of the target user.

[0232] In one example, the second feature similarity acquisition unit is specifically used to acquire the second feature similarity between the target user and the highly active user based on the user's second feature; the social relationship adjustment unit is specifically used to select a target highly active user from multiple highly active users as the target user's new social user based on the second feature similarity between the target user and the highly active user.

[0233] In one example, the apparatus may further include: a representative high-activity user acquisition unit, configured to acquire representative high-activity users of multiple high-activity user interest categories, wherein the representative high-activity users are determined based on the similarity between historical interaction data objects of high-activity users and data objects corresponding to the interest categories of high-activity users; a second feature similarity acquisition unit, specifically configured to acquire the second feature similarity between the target user and the representative high-activity users based on the second feature of the user; and a social relationship adjustment unit, specifically configured to select a target representative high-activity user from the representative high-activity users of the multiple high-activity user interest categories based on the second feature similarity between the target user and the representative high-activity users.

[0234] In one example, the apparatus may further include: a high-activity user interest class acquisition unit, used to cluster historical interaction data objects of high-activity users to form the plurality of high-activity user interest classes.

[0235] In one example, the apparatus may further include: a representative high-activity user acquisition unit, specifically configured to: obtain the data object correlation between the high-activity user and the high-activity user interest class based on the historical interaction data objects of the high-activity user and the data objects corresponding to the high-activity user interest class; obtain the high-activity user interest class corresponding to the high-activity user based on the data object correlation; and obtain representative high-activity users of the high-activity user interest class based on the data object correlation between the high-activity users associated with the high-activity user interest class and the high-activity user interest class.

[0236] In one example, the social relationship adjustment unit is specifically used to select a target representative high-active user from representative high-active users of the plurality of high-active user interest classes based on the increase ratio of high-active users and the similarity of the second feature.

[0237] In one example, the increase ratio of highly active users is determined as follows: the increase ratio is determined based on the number of data objects that the target user has interacted with.

[0238] In one example, the second feature adjustment unit is specifically used to obtain a seventh feature of the target user being affected by the target high-activity user based on the second feature similarity between the target user and the target high-activity user, and the second feature of the target high-activity user; to obtain an eighth feature of the target user based on the second feature and the seventh feature; and to obtain the adjusted second feature based on the eighth feature.

[0239] In one example, the second feature adjustment unit is specifically used to obtain the sixth feature of the target user based on the first feature of the target user; and to obtain the adjusted second feature based on the sixth feature, the second feature, and the eighth feature of the target user.

[0240] In one example, the apparatus may further include: a representative high-activity user acquisition unit, configured to acquire representative high-activity users of multiple high-activity user interest categories; a second feature similarity acquisition unit, specifically configured to acquire, based on the second feature of the user, a second feature similarity between the target user and social users, and a second feature similarity between the target user and the representative high-activity users; and a social relationship adjustment unit, specifically configured to, based on the second feature similarity between the target user and social users, determine noisy users among the social users and delete the noisy users from the social users; and, based on the second feature similarity between the target user and the representative high-activity users, select a target high-activity user from the representative high-activity users of multiple high-activity user interest categories. The target user is identified as a newly added social user; the second feature adjustment unit is specifically used to obtain a fourth feature of the target user influenced by the remaining social users based on the second feature similarity between the target user and the remaining social users after deleting noise users, and the second or third feature of the remaining social users; to obtain a seventh feature of the target user influenced by the newly added neighboring users based on the second feature similarity between the target user and the newly added neighboring users, and the second feature of the newly added neighboring users; to obtain a ninth feature of the target user based on the second feature, the fourth feature, and the seventh feature; and to obtain the adjusted second feature based on the ninth feature.

[0241] In one example, the second feature adjustment unit is specifically used to obtain the sixth feature of the target user based on the first feature of the target user; and to obtain the adjusted second feature based on the sixth feature, the second feature, and the ninth feature of the target user.

[0242] In one example, the process of obtaining the second feature similarity between users based on the user's second feature is performed multiple times; the social relationship information is adjusted based on the second feature similarity; and the second feature of the target user is adjusted based on the adjusted social users.

[0243] In one example, the second feature acquisition unit is a feature encoder included in the social recommendation model, the second feature similarity acquisition unit and the social relationship adjustment unit are social graph adjusters included in the social recommendation model, the second feature adjustment unit is a feature adjuster included in the social recommendation model, and the data object selection unit is a data object selector included in the social recommendation model. Accordingly, the feature encoder included in the social recommendation model acquires the second features of the user and the second features of the data object based on the interaction relationship information, the user's first feature, and the data object's first feature; the social graph adjuster included in the social recommendation model acquires the second feature similarity between users based on the user's second feature; the social relationship information is adjusted based on the second feature similarity; the feature adjuster included in the social recommendation model adjusts the second feature of the target user based on the adjusted social relationship information; and the data object selector included in the social recommendation model determines the data object to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0244] Third Embodiment

[0245] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push method. This method corresponds to Embodiment 1 of the above method. Since this method embodiment is basically similar to Method Embodiment 1, it is described simply. For relevant details, please refer to the description of Method Embodiment 1. The method embodiments described below are merely illustrative.

[0246] This application also provides a method for pushing data objects, including:

[0247] Step 1: Receive the data object recommendation information retrieval request sent by the client.

[0248] Clients include, but are not limited to, mobile communication devices, namely, mobile phones or smartphones, as well as personal computers, tablets, iPads, and other terminal devices. The request may carry the identifier of the target user.

[0249] Step 2: Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics.

[0250] Step 3: Based on the interaction relationship information, the user's first feature, and the data object's first feature, obtain the user's second feature and the data object's second feature.

[0251] Step 4: Based on the user's second feature, obtain the second feature similarity between users.

[0252] Step 5: Adjust the social relationship information based on the second feature similarity.

[0253] Step 6: Adjust the second characteristic of the target user based on the adjusted social relationship information.

[0254] Step 7: Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, determine the data object to be recommended to the target user.

[0255] Step 8: Provide data object recommendation information to the client.

[0256] The client can display recommended information about data objects.

[0257] In specific implementation, step 8 can be implemented as follows: If the target user is a low-activity user, then a webpage displaying the recommended data objects in a first layout is provided to the client; if the target user is a high-activity user, then a webpage displaying the recommended data objects in a second layout is provided to the client. The first layout can be arranging the multiple recommended data objects horizontally, and the second layout can be arranging the multiple recommended data objects vertically. This approach allows for different layouts of the recommended data objects for low-activity and high-activity users, improving the display effect of the recommended data objects and thus enhancing the user experience.

[0258] As can be seen from the above embodiments, the data object push method provided in this application involves receiving a data object recommendation information acquisition request sent by a client; acquiring the social relationship information of the target user, as well as the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; acquiring the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; acquiring the second feature similarity between users based on the user's second feature; adjusting the social relationship information based on the second feature similarity; adjusting the target user's second feature based on the adjusted social relationship information; determining the data object to recommend to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature; and providing the data object information to the client. This processing method introduces social relationship adjustment into social recommendation, improving the target user's social relationships, such as deleting noisy neighbors from the target user's existing social relationships and strengthening the connection between the target user and highly active users, thereby improving the quality of the user's social relationships and ultimately improving the quality of user understanding; therefore, it can effectively improve the social recommendation effect.

[0259] Fourth embodiment

[0260] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0261] This application also provides a data object push device, including: a request receiving unit, a data acquisition unit, a second feature acquisition unit, a second feature similarity acquisition unit, a social relationship adjustment unit, a second feature adjustment unit, a data object selection unit, and a recommendation information sending unit.

[0262] The system includes: a request receiving unit for receiving a data object recommendation information acquisition request sent by a client; a data acquisition unit for acquiring social relationship information of a target user, interaction relationship information between the user and the data object, a first feature of the user, and a first feature of the data object; a second feature acquisition unit for acquiring a second feature of the user and a second feature of the data object based on the interaction relationship information, the first feature of the user, and the first feature of the data object; a second feature similarity acquisition unit for acquiring the second feature similarity between users based on the second feature of the user; a social relationship adjustment unit for adjusting the social relationship information based on the second feature similarity; a second feature adjustment unit for adjusting the second feature of the target user based on the adjusted social relationship information; a data object selection unit for determining the data object to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object; and a recommendation information sending unit for providing data object recommendation information to the client.

[0263] Fifth embodiment

[0264] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push method. This method corresponds to Embodiment 1 of the above method. Since this method embodiment is basically similar to Method Embodiment 1, it is described simply. For relevant details, please refer to the description of Method Embodiment 1. The method embodiments described below are merely illustrative.

[0265] This application also provides a method for pushing data objects, including:

[0266] Step 1: Send a request to the server to obtain data object recommendation information;

[0267] Step 2: Receive data object recommendation information sent back by the server; the data object recommendation information is determined by social recommendation based on the adjusted social relationship information of the target user, the adjusted social relationship information is determined based on the second feature similarity between the target user's historical interaction data objects and other users, and the user's second feature is determined based on the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature;

[0268] Step 3: Display the recommended information for the data object.

[0269] In specific implementation, step 3 can be implemented in the following way: if the target user is a low-activity user, the data object recommendation information is displayed in the first layout; if the target user is a high-activity user, the data object recommendation information is displayed in the second layout.

[0270] As can be seen from the above embodiments, the data object push method provided in this application involves sending a data object recommendation information acquisition request to the server; receiving data object recommendation information returned by the server; the data object recommendation information is determined based on the adjusted social relationship information of the target user, the adjusted social relationship information is determined based on the second feature similarity between the target user's historical interaction data objects and other users, the user's second feature is determined based on the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; and displaying the data object recommendation information. This processing method allows the client to obtain data object recommendation information determined based on the target user's improved social relationships. Because the quality of the user's social relationships is improved, the quality of user understanding is also improved; therefore, the social recommendation effect can be effectively improved.

[0271] Sixth Embodiment

[0272] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0273] This application also provides a data object push device, including: a request sending unit, a display receiving unit, and a recommendation information display unit.

[0274] The system includes a request sending unit for sending a request to the server to obtain data object recommendation information; a display receiving unit for receiving data object recommendation information returned by the server; the data object recommendation information is determined based on the adjusted social relationship information of the target user, which is determined based on the second feature similarity between the target user's historical interaction data objects and other users, and the user's second feature is determined based on the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; and a recommendation information display unit for displaying the data object recommendation information.

[0275] Seventh Embodiment

[0276] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push method. This method corresponds to Embodiment 1 of the above method. Since this method embodiment is basically similar to Method Embodiment 1, it is described simply. For relevant details, please refer to the description of Method Embodiment 1. The method embodiments described below are merely illustrative.

[0277] This application also provides a method for pushing data objects, including:

[0278] Step 1: Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics;

[0279] Step 2: Based on the interaction relationship information, the user's first feature, and the data object's first feature, obtain the user's second feature and the data object's second feature;

[0280] Step 3: Based on the second feature of the target user and the second feature of the social user, obtain the second feature similarity between the target user and the social user;

[0281] Step 4: Based on the second feature similarity, identify the noisy users among the social users;

[0282] Step 5: Remove the noisy user from the social users list;

[0283] Step 6: Adjust the second characteristic of the target user based on the adjusted social relationship information of the target user;

[0284] Step 7: Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, determine the data object to be recommended to the target user.

[0285] As can be seen from the above embodiments, the data object push method provided in this application obtains the social relationship information of the target user, as well as the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; obtains the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; obtains the second feature similarity between the target user and the social user based on the target user's second feature and the social user's second feature; identifies noisy users among the social users based on the second feature similarity; deletes the noisy users from the social users; adjusts the target user's second feature based on the target user's adjusted social relationship information; and determines the data object to be recommended to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature. This processing method introduces social relationship adjustment into social recommendation, improving the quality of the user's social relationships by deleting noisy neighbors in the target user's (e.g., low-activity user) original social relationships, thereby improving the quality of user understanding; therefore, it can effectively improve the social recommendation effect.

[0286] Eighth embodiment

[0287] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0288] This application also provides a data object push device, including: a data acquisition unit, a second feature acquisition unit, a second feature similarity acquisition unit, a noisy user determination unit, a social relationship adjustment unit, a second feature adjustment unit, and a data object selection unit.

[0289] A data acquisition unit is used to acquire the social relationship information of a target user, as well as the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; a second feature acquisition unit is used to acquire the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; a second feature similarity acquisition unit is used to acquire the second feature similarity between the target user and the social user based on the target user's second feature and the social user's second feature; a noisy user determination unit is used to determine noisy users among the social users based on the second feature similarity; a social relationship adjustment unit is used to delete the noisy users from the social users; a second feature adjustment unit is used to adjust the target user's second feature based on the target user's adjusted social relationship information; and a data object selection unit is used to determine the data object to be recommended to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature.

[0290] Ninth Embodiment

[0291] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push method. This method corresponds to Embodiment 1 of the above method. Since this method embodiment is basically similar to Method Embodiment 1, it is described simply. For relevant details, please refer to the description of Method Embodiment 1. The method embodiments described below are merely illustrative.

[0292] This application also provides a method for pushing data objects, including:

[0293] Step 1: Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics;

[0294] Step 2: Based on the interaction relationship information, the user's first feature, and the data object's first feature, obtain the user's second feature and the data object's second feature;

[0295] Step 3: Obtain representative high-active users based on their interest categories;

[0296] Step 4: Based on the second feature of the target user and the second feature of the representative highly active user, obtain the second feature similarity between the target user and the representative highly active user;

[0297] Step 5: Based on the second feature similarity between the target user and the representative high-activity user, select the target representative high-activity user from the representative high-activity users of the multiple high-activity user interest categories, and use them as the target user's new social users;

[0298] Step 6: Adjust the second characteristic of the target user based on the adjusted social relationship information of the target user;

[0299] Step 7: Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, determine the data object to be recommended to the target user.

[0300] As can be seen from the above embodiments, the data object push method provided in this application obtains the social relationship information of the target user, the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature; obtains the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; obtains representative high-active users of multiple high-active user interest categories; obtains the second feature similarity between the target user and the representative high-active users based on the target user's second feature and the representative high-active users' second features; selects a target representative high-active user from the representative high-active users of the multiple high-active user interest categories based on the second feature similarity between the target user and the representative high-active users, as the target user's new social user; adjusts the target user's second feature based on the target user's adjusted social relationship information; and determines the data object to be recommended to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature. This approach incorporates social relationship adjustments into social recommendations. By strengthening the connection between target users (such as low-activity users) and high-activity users, it increases the quantity of users' social relationships while ensuring the quality of newly added relationships, thereby improving the quality of user understanding. Therefore, it can effectively enhance the performance of social recommendations. Simultaneously, by creating interest categories for high-activity users, mining their data object preferences, and identifying representative high-activity users within these interest categories, it strengthens the connection between target users and representative high-activity users. This allows for efficient guidance of the connection strengthening process between target users and high-activity users, effectively reducing computational complexity and improving the efficiency of improving user social relationships, thus enhancing the efficiency of social recommendations.

[0301] Tenth Embodiment

[0302] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0303] This application also provides a data object push device, including: a data acquisition unit, a second feature acquisition unit, a representative high-activity user acquisition unit, a second feature similarity acquisition unit, a social relationship adjustment unit, a second feature adjustment unit, and a data object selection unit.

[0304] The system includes the following components: a data acquisition unit for acquiring social relationship information of a target user, interaction relationship information between the user and a data object, a first feature of the user, and a first feature of the data object; a second feature acquisition unit for acquiring a second feature of the user and a second feature of the data object based on the interaction relationship information, the first feature of the user, and the first feature of the data object; a representative high-activity user acquisition unit for acquiring representative high-activity users from multiple high-activity user interest categories; a second feature similarity acquisition unit for acquiring the second feature similarity between the target user and the representative high-activity users based on the second feature similarity between the target user and the representative high-activity users; a social relationship adjustment unit for selecting a target representative high-activity user from the representative high-activity users of the multiple high-activity user interest categories as a new social user for the target user based on the second feature similarity between the target user and the representative high-activity users; a second feature adjustment unit for adjusting the second feature of the target user based on the adjusted social relationship information of the target user; and a data object selection unit for determining the data object to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0305] Eleventh Embodiment

[0306] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push method. This method corresponds to Embodiment 1 of the above method. Since this method embodiment is basically similar to Method Embodiment 1, it is described simply. For relevant details, please refer to the description of Method Embodiment 1. The method embodiments described below are merely illustrative.

[0307] This application also provides a method for pushing data objects, including:

[0308] Step 1: Obtain the target user's social relationship information.

[0309] Step 2: Based on the historical interaction data objects of the social users and the historical interaction data objects of the target users, obtain the data object similarity between the social users and the target users.

[0310] Step 3: Based on the similarity of the data objects, identify the noisy users among the social users.

[0311] Step 4: Based on social users other than the noise users, determine the data objects to recommend to the target users.

[0312] As can be seen from the above embodiments, the data object push method provided in this application obtains the social relationship information of the target user; obtains the similarity between the social user and the target user based on the historical interaction data objects of the social user and the historical interaction data objects of the target user; identifies noisy users among the social users based on the similarity; and determines the data objects to be recommended to the target user based on the remaining social users other than the noisy users. This processing method introduces social relationship adjustment into social recommendation, improving the quality of the user's social relationships and thus the quality of user understanding by deleting noisy neighbors in the original social relationships of the target user (such as a low-activity user); therefore, it can effectively improve the social recommendation effect.

[0313] Twelfth Embodiment

[0314] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0315] This application also provides a data object push device, including: a social relationship information acquisition unit, a data object similarity acquisition unit, a noisy user identification unit, and a data object recommendation information identification unit.

[0316] The system includes a social relationship information acquisition unit for acquiring social relationship information of a target user; a data object similarity acquisition unit for acquiring data object similarity between a social user and a target user based on the historical interaction data objects of the social user and the historical interaction data objects of the target user; a noise user identification unit for identifying noise users among the social users based on the data object similarity; and a data object recommendation information identification unit for determining data objects to recommend to the target user based on social users other than the noise users.

[0317] Thirteenth Embodiment

[0318] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push method. This method corresponds to Embodiment 1 of the above method. Since this method embodiment is basically similar to Method Embodiment 1, it is described simply. For relevant details, please refer to the description of Method Embodiment 1. The method embodiments described below are merely illustrative.

[0319] This application also provides a method for pushing data objects, including:

[0320] Step 1: Obtain multiple interest categories of highly active users;

[0321] Step 2: Based on the historical interaction data objects of highly active users and the data objects corresponding to the interest classes of highly active users, obtain the correlation between the data objects of highly active users and the interest classes of highly active users;

[0322] Step 3: Based on the relevance of the data objects, obtain the interest categories of the highly active users corresponding to the highly active users;

[0323] Step 4: Based on the data object correlation between the high-active users corresponding to the high-active user interest class and the high-active user interest class, obtain the representative high-active users of the high-active user interest class;

[0324] Step 5: Based on the historical interaction data objects of the representative highly active users and the historical interaction data objects of the target users, obtain the similarity between the representative highly active users and the target users;

[0325] Step 6: Based on the similarity, select representative high-active users related to the target user from the representative high-active users of the multiple high-active user interest categories;

[0326] Step 7: Based on the relevant representative highly active users, determine the data objects to be recommended to the target user.

[0327] As can be seen from the above embodiments, the data object push method provided in this application involves: acquiring multiple high-activity user interest categories; obtaining the data object relevance between the high-activity user and the high-activity user interest category based on the historical interaction data objects of the high-activity user and the data objects corresponding to the high-activity user interest category; acquiring the high-activity user interest category corresponding to the high-activity user based on the data object relevance; acquiring representative high-activity users of the high-activity user interest category based on the data object relevance between the high-activity user corresponding to the high-activity user interest category and the high-activity user interest category; obtaining the similarity between the representative high-activity user and the target user based on the historical interaction data objects of the representative high-activity user and the target user; selecting representative high-activity users related to the target user from the representative high-activity users of the multiple high-activity user interest categories based on the similarity; and determining the data object to be recommended to the target user based on the related high-activity users. This approach incorporates social relationship adjustments into social recommendations. By strengthening the connection between target users (such as low-activity users) and high-activity users, it increases the quantity of users' social relationships while ensuring the quality of newly added relationships, thereby improving the quality of user understanding. Therefore, it can effectively enhance the performance of social recommendations. Simultaneously, by creating interest categories for high-activity users, mining their data object preferences, and identifying representative high-activity users within these interest categories, it strengthens the connection between target users and representative high-activity users. This allows for efficient guidance of the connection strengthening process between target users and high-activity users, effectively reducing computational complexity and improving the efficiency of improving user social relationships, thus enhancing the efficiency of social recommendations.

[0328] Fourteenth Embodiment

[0329] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a data object push device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0330] This application also provides a data object push device, including: a high-activity user interest category acquisition unit, a data object relevance acquisition unit, a high-activity user classification unit, a representative high-activity user acquisition unit, a similarity acquisition unit, a representative high-activity user selection unit, and a data object recommendation information determination unit.

[0331] The system includes the following components: a high-activity user interest category acquisition unit, used to acquire multiple high-activity user interest categories; a data object relevance acquisition unit, used to acquire the data object relevance between a high-activity user and the high-activity user interest category based on the historical interaction data objects of the high-activity user and the data objects corresponding to the high-activity user interest category; a high-activity user classification unit, used to acquire the high-activity user interest category corresponding to the high-activity user based on the data object relevance; a representative high-activity user acquisition unit, used to acquire representative high-activity users of the high-activity user interest category based on the data object relevance between the high-activity users corresponding to the high-activity user interest category and the high-activity user interest category; a similarity acquisition unit, used to acquire the similarity between the representative high-activity user and the target user based on the historical interaction data objects of the representative high-activity user and the historical interaction data objects of the target user; a representative high-activity user selection unit, used to select representative high-activity users related to the target user from the representative high-activity users of the multiple high-activity user interest categories based on the similarity; and a data object recommendation information determination unit, used to determine the data objects to be recommended to the target user based on the related representative high-activity users.

[0332] Fifteenth Embodiment

[0333] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides a social recommendation model processing method. This method corresponds to Embodiment 1 of the above method. Since this method embodiment is basically similar to Method Embodiment 1, it is described simply. For relevant details, please refer to the description of Method Embodiment 1. The method embodiments described below are merely illustrative.

[0334] This application also provides a social recommendation model processing method, including:

[0335] Step 401: Obtain multiple training samples.

[0336] The training sample data includes users' social relationship information, interaction relationship information between users and data objects, users' first features, data objects' first features, and data object recommendation information. Additionally, the training sample data may also include users' second features. The training samples can be highly active users or inactive users, and the multiple training samples include training samples from highly active users and training samples from inactive users. The data from the highly active user training samples may also include the interest categories corresponding to the highly active users. In specific implementations, the training samples from inactive users can be used as positive samples, and the training samples from highly active users can be used as negative samples.

[0337] Step 403: Construct the network structure of the social recommendation model.

[0338] The network structure includes a feature encoder, a social graph adjuster, a feature adjuster, and a data object selector. The feature encoder is used to obtain a second feature of the user and a second feature of the data object based on the interaction relationship information, a first feature of the user, and a first feature of the data object. The social graph adjuster is used to obtain the second feature similarity between users based on the second feature of the user; and adjust the social relationship information based on the second feature similarity. The feature adjuster is used to adjust the second feature of the target user based on the adjusted social relationship information. The data object selector is used to determine the data objects to recommend to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

[0339] Step 405: Learn the parameters of the model from the multiple training samples.

[0340] The parameters can be r1, r2, α, etc. from Example 1. w f b f wait.

[0341] In this embodiment, supervised machine learning is employed to learn the model's parameters from the multiple training samples. Specifically, the social relationship information of the target user, the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature—all included in the training samples—can be used as input data for the model. The model outputs data object recommendation information, which is the model's predicted data. Based on the difference between the model's predicted data and the data object recommendation information included in the training samples, the model parameters are adjusted until the difference between the model's predicted data and the data object recommendation information included in the training samples reaches the optimization target.

[0342] In one example, the social graph adjuster includes a module for adding highly active users, used to obtain representative highly active users of multiple highly active user interest categories; obtain the similarity between the target user and the representative highly active users based on the second feature of the target user and the second feature of the representative highly active users; and select a target representative highly active user from the representative highly active users of the multiple highly active user interest categories as the target user's new social user based on the similarity between the target user and the representative highly active user.

[0343] In the process of selecting representative high-active users from those with similar second characteristics to the target user from among the representative high-active users' interest categories, the target user's second characteristic is used. However, if the target user is a low-active user, the above processing on low-active users will not reflect their true interests and preferences due to the insufficient quality of their representations. This will affect the effect of increasing the number of high-active users, fail to support an effective matching process between low-active and high-active users, and make it difficult to match low-active users with suitable high-active user interest categories.

[0344] In one example, the method provided in this application embodiment may further include the following steps:

[0345] Step S501: For the training samples of highly active users, adjust the second feature of the highly active users according to the feature distribution of low-active users, and treat the highly active users as pseudo-low-active users.

[0346] The training samples can include training samples from highly active users and those from less active users. Based on the training samples from less active users, the feature distribution of these users can be obtained. For the training samples from highly active users, the second feature of the highly active users is adjusted according to the feature distribution of the less active users. Since adjusting the second feature of highly active users makes their feature distribution similar to that of less active users, these highly active users with adjusted second features are called pseudo-less active users.

[0347] In one example, adjusting the second feature of a highly active user based on the feature distribution of inactive users can be achieved by randomly reducing the amount of information in the second feature of the highly active user. This method is called Random Masking. The essence of mimicking inactive users is to reduce... The amount of information contained within. Therefore, a natural approach is to randomly omit it. A portion of the vector can be replaced with 0. In practice, a random vector can be sampled first. Where d is Dimensions. Each digit is independently sampled from a Bernoulli distribution. The modified second feature of highly active users can be represented as:

[0348] Random mask: Where 'o' indicates element-wise multiplication.

[0349] In one example, adjusting the second feature of highly active users based on the feature distribution of low-active users can be achieved as follows: Adjust the second feature of highly active users based on the feature distributions of both highly active and low-active users. This method is called Distribution Shift, which focuses on modeling the feature distributions of highly active and low-active users and transforming the former into the latter, making the second feature distribution of highly active users the same as or similar to that of low-active users. In a training round, there is a group of highly active users S. + and a group of low-activity users S - .

[0350] In specific implementation, the second feature of high-active users is adjusted based on the feature distribution of high-active users and low-active users. This can be achieved in the following way: 1) Obtain the second features of multiple high-active users and multiple low-active users; 2) Based on the second features of the multiple high-active users, obtain the first mean and first deviation of the second features of the multiple high-active users; based on the second features of the multiple low-active users, obtain the second mean and second deviation of the second features of the multiple low-active users; 3) Based on the first mean, the second features of the high-active users, the second mean, and the second deviation, obtain the third deviation of the high-active users; 4) Based on the first mean and the third deviation, obtain the adjusted second feature of the high-active users. In specific implementation, the following methods can be used to change...

[0351] Distribution drift: Where, μ(S) and These are used to calculate the mean and standard deviation of set S, respectively.

[0352] In one example, adjusting the second feature of a highly active user based on the feature distribution of inactive users can be achieved as follows: the second feature of the highly active user is adjusted to be a weighted sum of the second features of the inactive user and the second feature of the highly active user. This method is called Inactive Mixture. In recommendation scenarios, by injecting inactive information (positive samples) into the features of highly active users (negative samples), more difficult negative samples are generated, thereby effectively improving the effect of supervised learning.

[0353] In specific implementation, the weighted sum of the second characteristics of low-activity users and high-activity users is used as the adjusted second characteristic of high-activity users. This can be achieved in the following way: For each high-activity user u + One or more low-activity users are randomly selected, and the weighted sum of the second features of the low-activity users and the second features of the high-activity users is used as the adjusted second feature of the high-activity users. In specific implementation, the following formula can be used for calculation:

[0354] Low-activity mixture:

[0355] Where e u- It is the second feature of a selected low-activity user, with a weight β∈(0,1).

[0356] Step S503: The interest class of the highly active user corresponding to the highly active user is used as the interest class of the highly active user corresponding to the pseudo-low-activity user. Table 4 shows the correspondence between the pseudo-low-activity user and the interest class of the highly active user in this embodiment.

[0357]

[0358] Table 4. Correspondence between interest categories of pseudo-low-activity users and high-activity users

[0359] In this case, step 405 can be implemented as follows: the parameters of the model are learned from multiple second training samples. The multiple second training samples include training samples from low-activity users and training samples from pseudo-low-activity users. The data of the training samples from pseudo-low-activity users includes the second features of the pseudo-low-activity users, i.e., the adjusted second features of the high-activity users.

[0360] If the second feature of highly active users is not adjusted, step 405 can be implemented as follows: learn the parameters of the model from multiple first training samples. The multiple first training samples include training samples of low-active users and training samples of highly active users, and the data of the training samples of highly active users includes the second feature of highly active users.

[0361] The method provided in this application embodiment, through the above steps S501 to S503, slightly changes the high-activity user u + The representation of By obfuscating the characteristics of highly active and inactive users, pseudo-inactive high-active users can "mimic" inactive users, and the altered characteristics of users... + It can still belong to the original high-activity user interest category C(u) + Using "pseudo-low-activity high-activity users" as training samples, the model achieves imitation learning based on feature confusion, and can train the model to learn how to find suitable high-activity users as new social friends for low-activity users.

[0362] Let Θ denote the parameters of the feature encoder in the social recommendation model, and Ω denote the parameters of the social graph adjuster. To effectively train these two sets of parameters, in one example, the following iterative optimization steps can be used:

[0363] Step S601: Keep the parameters of the social graph adjuster unchanged, and learn the parameters of the feature encoder from the plurality of first training samples.

[0364] Freeze Ω, train Θ, continue T Θ Round. In this stage, the structure of the social graph is fixed, and the user's adjusted second feature contains both the collaborative information described in Example 1 and the original social information. In specific implementation, the social recommendation model can be optimized using Bayesian Personalized Ranking (BPR) loss:

[0365]

[0366] Each ternary instance in θ consists of user u, a data object i that u has purchased, and an arbitrarily selected data object i'. ui Indicates whether user u purchased data object i, r ui′ This indicates whether user u has selected data object i', in reality r ui =1 and r ui′ =0. It is the predicted probability that a user will purchase data object i. This represents the predicted probability of the user-selected data object i'. The λ-weighted L2 norm reduces the generalization error.

[0367] Step S603: Keeping the parameters of the feature encoder unchanged, learn the parameters of the social graph adjuster from the plurality of second training samples.

[0368] Freeze Θ, train Ω, continue T Ω The goal of this stage is to achieve a better social graph structure.

[0369] In one example, the model loss function used when training the model in step S603 may include a social recommendation loss term and a mimicry loss term. The mimicry loss term is determined based on a first similarity and a second similarity. The first similarity includes the second feature similarity between the pseudo-low-activity user and a representative high-activity user of the high-activity user interest class corresponding to the pseudo-low-activity user. The second similarity includes the second feature similarity between the pseudo-low-activity user and representative high-activity users of other high-activity user interest classes. Specifically, the mimicry loss term can be calculated using the following formula:

[0370] Loss of imitation:

[0371] in, It is a pseudo-low-activity user u + The second characteristic, as described above Assume C(u) + ) = c j This indicates a fake high-activity user u +It belongs to the j-th high-activity user interest class. Ψ is the cosine similarity, and τ is the temperature coefficient. The second characteristic represents the representative high-active user's interest category, corresponding to pseudo-low-active users. This represents the second characteristic of a representative high-active user, other than the high-active user interest class corresponding to pseudo-low-active users.

[0372] The optimization objectives are as follows:

[0373] Overall loss: L Ω =L bpr (Ω)+ξL ml

[0374] Where ξ is the combination coefficient.

[0375] By adopting the above processing method, the model is trained by comparing and contrasting the similarity between pseudo-low-activity users and their corresponding high-activity users' interest categories. This allows the model to be trained more accurately in associating low-activity users with high-activity users.

[0376] To effectively train the social recommendation model, an iterative optimization strategy is used to train the feature encoder and social graph adjuster, iteratively performing the two optimization steps described above. The entire model can then be backpropagated using stochastic gradient descent to update the model parameters.

[0377] As can be seen from the above embodiments, the data object push method provided in this application obtains multiple training samples; the training samples include the social relationship information of the target user, the interaction relationship information between the user and the data object, the user's first feature, the data object's first feature, and the data object recommendation information; constructs a network structure for a social recommendation model; the network structure includes a feature encoder, a social graph adjuster, a feature adjuster, and a data object selector; the feature encoder is used to obtain the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; the social graph adjuster is used to obtain the second feature similarity between users based on the user's second feature; adjusts the social relationship information based on the second feature similarity; the feature adjuster is used to adjust the target user's second feature based on the adjusted social relationship information; the data object selector is used to determine the data object to be recommended to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature; and learns the model's parameters from the multiple training samples. This approach allows for the introduction of a social graph adjuster into the social recommendation model. By adjusting the social graph, the model can improve the social relationships of target users (such as low-activity users), for example, by removing noisy neighbors from the target user's existing social relationships and strengthening the connection between the target user and high-activity users. This improves the quality of user social relationships and, consequently, the quality of user understanding. Therefore, it can effectively improve the quality of the social recommendation model and thus enhance the effectiveness of social recommendations.

[0378] Sixteenth Embodiment

[0379] In the above embodiments, a social recommendation model processing method is provided. Correspondingly, this application also provides a social recommendation model processing apparatus. This apparatus corresponds to the embodiments of the above method. Since the apparatus embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant parts can be referred to in the description of the method embodiments. The apparatus embodiments described below are merely illustrative.

[0380] This application also provides a social recommendation model processing device, including: a training sample acquisition unit, a network structure construction unit, and a parameter adjustment unit.

[0381] A training sample acquisition unit is used to acquire multiple training samples; the training samples include user social relationship information, interaction relationship information between users and data objects, user's first feature, data object's first feature, and data object recommendation information; a network structure construction unit is used to construct the network structure of the social recommendation model; the network structure includes a feature encoder, a social graph adjuster, a feature adjuster, and a data object selector; the feature encoder is used to acquire the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature; the social graph adjuster is used to acquire the second feature similarity between users based on the user's second feature; and adjust the social relationship information based on the second feature similarity; the feature adjuster is used to adjust the target user's second feature based on the adjusted social relationship information; the data object selector is used to determine the data object to recommend to the target user based on the similarity between the target user's adjusted second feature and the data object's second feature; and a parameter adjustment unit is used to learn the model's parameters from the multiple training samples.

[0382] In one example, the social graph adjuster includes a module for adding highly active users, used to acquire representative highly active users of multiple highly active user interest categories; to acquire the similarity between the target user and the representative highly active users based on the second feature of the target user and the second feature of the representative highly active users; and to select a target representative highly active user from the representative highly active users of the multiple highly active user interest categories based on the similarity between the target user and the representative highly active users, as a new social user of the target user. Correspondingly, the device may further include: a second feature adjustment unit, used to adjust the second feature of the highly active user based on the feature distribution of low-active users for the training samples of highly active users, thus treating the highly active user as a pseudo-low-active user; and a highly active user interest category acquisition unit, used to obtain the highly active user interest category corresponding to the highly active user. The corresponding pseudo-low-activity user corresponds to the high-activity user interest class; correspondingly, the parameter adjustment unit is used to learn the parameters of the feature encoder from multiple first training samples, the multiple first training samples including training samples of high-activity users; according to the model loss function, the parameters of the social graph adjuster are learned from multiple second training samples, the multiple second training samples including training samples of pseudo-low-activity users; wherein, the model loss function includes a social recommendation loss term and an imitation loss term; the imitation loss term is determined according to a first similarity and a second similarity, the first similarity including the second feature similarity between the pseudo-low-activity user and the representative high-activity user of the high-activity user interest class corresponding to the pseudo-low-activity user, and the second similarity including the second feature similarity between the pseudo-low-activity user and the representative high-activity user of other high-activity user interest classes.

[0383] In one example, the second feature adjustment unit is specifically used to randomly reduce the amount of information in the second feature of highly active users; or, adjust the second feature of highly active users according to the feature distribution of highly active users and the feature distribution of low-active users; or, adjust the second feature of highly active users to be a weighted sum of the second feature of low-active users and the second feature of highly active users.

[0384] In one example, adjusting the second feature of a high-active user based on the feature distribution of high-active users and the feature distribution of low-active users includes: obtaining the second features of multiple high-active users and multiple low-active users; obtaining a first mean and a first deviation of the second features of the multiple high-active users based on the second features of the multiple high-active users; obtaining a second mean and a second deviation of the second features of the multiple low-active users based on the second features of the multiple low-active users; obtaining a third deviation of the high-active users based on the first mean, the second features of the high-active users, the second mean, and the second deviation; and obtaining the adjusted second feature of the high-active users based on the first mean and the third deviation.

[0385] Seventeenth Embodiment

[0386] In the above embodiments, a data object push method is provided. Correspondingly, this application also provides an electronic device. This device corresponds to the embodiments of the above method. Since the device embodiments are basically similar to the method embodiments, the description is relatively simple, and relevant details can be found in the description of the method embodiments. The device embodiments described below are merely illustrative.

[0387] The electronic device of this embodiment includes: a processor; and a memory for storing a program for implementing the method described in any of the above embodiments, wherein the device is powered on and the program of the method is executed by the processor.

[0388] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties. Furthermore, the collection, use and processing of the relevant data must comply with the relevant laws, regulations and standards of the relevant countries and regions, and corresponding operation portals are provided for users to choose to authorize or refuse.

[0389] Although this application discloses preferred embodiments as described above, it is not intended to limit this application. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of this application. Therefore, the scope of protection of this application should be determined by the scope defined in the claims of this application.

[0390] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0391] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0392] 1. Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information by any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include non-transitory computer-readable media, such as modulated data signals and carrier waves.

[0393] 2. Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

Claims

1. A data object push method, characterized by, include: Obtain the target user's social relationship information, the interaction relationship information between the user and the data object, the user's primary characteristics, and the primary characteristics of the data object; The interaction relationship information includes: a set of data objects that interact with the user, and a set of users that interact with the data objects; Based on the interaction relationship information, the user's first characteristic, and the data object's first characteristic, obtain the user's second characteristic and the data object's second characteristic; including: obtaining the user's second characteristic based on the set of data objects that interact with the user and the data object's first characteristic; obtaining the data object's second characteristic based on the set of users that interact with the data object and the user's first characteristic; Based on the user's second feature, obtain the second feature similarity between users; The social relationship information is adjusted based on the second feature similarity. The second characteristic of the target user is adjusted based on the adjusted social relationship information; Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object is determined to be recommended to the target user.

2. The method according to claim 1, characterized in that, The step of obtaining the second feature similarity between users based on the user's second feature includes: Based on the user's second feature, obtain the second feature similarity between the target user and the social user; The step of adjusting the social relationship information based on the second feature similarity includes: Based on the second feature similarity between the target user and the social users, noisy users among the social users are identified; Remove the noisy user from the social media user list.

3. The method according to claim 1, characterized in that, The step of obtaining the second feature similarity between users based on the user's second feature includes: Based on the user's second feature, obtain the second feature similarity between the target user and the highly active user; The step of adjusting the social relationship information based on the second feature similarity includes: Based on the second feature similarity between the target user and the highly active user, a target highly active user is selected from multiple highly active users as the target user's new social users.

4. The method of claim 3, wherein, Also includes: Obtain representative high-active users across multiple high-active user interest categories; The step of obtaining the second feature similarity between the target user and the highly active user based on the user's second feature includes: Based on the user's second feature, obtain the second feature similarity between the target user and the representative highly active user; The step of selecting a target high-activity user from multiple high-activity users based on the second feature similarity between the target user and the high-activity user includes: Based on the second feature similarity between the target user and the representative high-activity user, a target representative high-activity user is selected from the representative high-activity users of the multiple high-activity user interest categories.

5. The method according to claim 1, characterized in that, Also includes: Obtain representative high-active users across multiple high-active user interest categories; The step of obtaining the second feature similarity between users based on the user's second feature includes: Based on the user's second feature, obtain the second feature similarity between the target user and the social user, and the second feature similarity between the target user and the representative highly active user; The step of adjusting the social relationship information based on the second feature similarity includes: Based on the second feature similarity between the target user and the social users, noisy users are identified among the social users and then removed from the social users. Based on the second feature similarity between the target user and the representative high-activity user, a target representative high-activity user is selected from representative high-activity users of multiple high-activity user interest categories as the target user's new social users; The step of adjusting the second characteristic of the target user based on the adjusted social user includes: Based on the second feature similarity between the target user and the remaining social users after removing noisy users, and the second or third feature of the remaining social users, a fourth feature of the target user's influence by the remaining social users is obtained; Based on the second feature similarity between the target user and the newly added neighboring user, and the second feature of the newly added neighboring user, a seventh feature is obtained to show how the target user is affected by the newly added neighboring user. Based on the second feature, the fourth feature, and the seventh feature of the target user, the ninth feature of the target user is obtained; Based on the ninth feature, the adjusted second feature is obtained.

6. The method according to claim 1, characterized in that, The social recommendation model uses a feature encoder to obtain the user's second feature and the data object's second feature based on the interaction relationship information, the user's first feature, and the data object's first feature. The social recommendation model includes a social graph adjuster, which obtains the second feature similarity between users based on the second feature of the user; and adjusts the social relationship information based on the second feature similarity. The second feature of the target user is adjusted based on the adjusted social relationship information using the feature adjuster included in the social recommendation model. The social recommendation model includes a data object selector, which determines the data objects to be recommended to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object.

7. A method for pushing data objects, characterized in that, include: Receive a request from the client to retrieve recommendation information for a data object; Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics; The interaction relationship information includes: a set of data objects that interact with the user, and a set of users that interact with the data objects; Based on the interaction relationship information, the user's first characteristic, and the data object's first characteristic, obtain the user's second characteristic and the data object's second characteristic; including: obtaining the user's second characteristic based on the set of data objects that interact with the user and the data object's first characteristic; obtaining the data object's second characteristic based on the set of users that interact with the data object and the user's first characteristic; Based on the user's second feature, obtain the second feature similarity between users; The social relationship information is adjusted based on the second feature similarity. The second characteristic of the target user is adjusted based on the adjusted social relationship information; Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object to be recommended to the target user is determined; Provide data object recommendation information to the client.

8. A method for pushing data objects, characterized in that, include: Send a request to the server to retrieve data object recommendation information; Receive data object recommendation information sent back by the server; The data object recommendation information is determined based on the adjusted social relationship information of the target user, including: obtaining the second feature similarity between users based on the user's second feature; adjusting the social relationship information based on the second feature similarity; adjusting the second feature of the target user based on the adjusted social relationship information; and determining the data object to recommend to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object. The adjusted social relationship information is determined based on the second feature similarity between the target user's historical interaction data objects and other users. The user's second feature is determined based on the interaction relationship information between the user and the data object, the user's first feature, and the data object's first feature. The interaction relationship information includes: a set of data objects that have interacted with the user, and a set of users that have interacted with the data objects. Displays recommendation information for the data object.

9. A method for pushing data objects, characterized in that, include: Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics; The interaction relationship information includes: a set of data objects that interact with the user, and a set of users that interact with the data objects; Based on the interaction relationship information, the user's first characteristic, and the data object's first characteristic, obtain the user's second characteristic and the data object's second characteristic; including: obtaining the user's second characteristic based on the set of data objects that interact with the user and the data object's first characteristic; obtaining the data object's second characteristic based on the set of users that interact with the data object and the user's first characteristic; Based on the second feature of the target user and the second feature of the social user, the second feature similarity between the target user and the social user is obtained; Based on the second feature similarity, noisy users among the social users are identified; Remove the noisy user from the social users; Adjust the second characteristic of the target user based on the adjusted social relationship information of the target user; Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object is determined to be recommended to the target user.

10. A method for pushing data objects, characterized in that, include: Obtain the target user's social relationship information, as well as the interaction relationship information between the user and the data object, the user's primary characteristics, and the data object's primary characteristics; The interaction relationship information includes: a set of data objects that interact with the user, and a set of users that interact with the data objects; Based on the interaction relationship information, the user's first characteristic, and the data object's first characteristic, obtain the user's second characteristic and the data object's second characteristic; including: obtaining the user's second characteristic based on the set of data objects that interact with the user and the data object's first characteristic; obtaining the data object's second characteristic based on the set of users that interact with the data object and the user's first characteristic; Obtain representative high-active users across multiple high-active user interest categories; Based on the second feature of the target user and the second feature of the representative high-activity user, obtain the second feature similarity between the target user and the representative high-activity user; Based on the second feature similarity between the target user and the representative high-activity user, a target representative high-activity user is selected from the representative high-activity users of the multiple high-activity user interest categories as the target user's new social users; Adjust the second characteristic of the target user based on the adjusted social relationship information of the target user; Based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object is determined to be recommended to the target user.

11. A method for pushing data objects, characterized in that, include: Acquire multiple interest categories of highly active users; The high-activity user interest category is a high-activity user category formed by clustering high-activity users according to their data object preferences; the data object is one of the following: product object, news object, video object, and advertising object; Based on the historical interaction data objects of highly active users and the data objects corresponding to the interest classes of highly active users, obtain the correlation between the data objects of highly active users and the interest classes of highly active users; Based on the relevance of the data objects, obtain the interest categories of the highly active users corresponding to the highly active users; Based on the correlation between the high-active users corresponding to the high-active user interest class and the data object of the high-active user interest class, obtain the representative high-active users of the high-active user interest class; Based on the historical interaction data objects of the representative highly active users and the historical interaction data objects of the target users, the similarity between the representative highly active users and the target users is obtained; Based on the similarity, representative high-active users related to the target user are selected from representative high-active users of the multiple high-active user interest categories; Based on the relevant representative highly active users, determine the data objects to be recommended to the target user, including: obtaining the second feature similarity between the target user and the representative highly active users based on the second feature of the target user and the second feature of the representative highly active users; Based on the second feature similarity between the target user and the representative highly active users, a target representative highly active user is selected from the representative highly active users of the multiple highly active user interest categories as the target user's new social users; based on the adjusted social relationship information of the target user, the second feature of the target user is adjusted; based on the similarity between the adjusted second feature of the target user and the second feature of the data object, a data object recommended to the target user is determined.

12. A method for pushing data objects, characterized in that, include: Obtain the target user's social relationship information; Based on the historical interaction data objects of social users and the historical interaction data objects of target users, the similarity of data objects between social users and target users is obtained; the data objects are one of the following: product objects, news objects, video objects, and advertising objects. Based on the similarity of the data objects, noisy users among the social users are identified; Based on social users other than the noisy users, determine the data objects to recommend to the target user, including: obtaining the second feature similarity between the target user and the social users based on the second feature of the target user and the second feature of the social users; identifying the noisy users among the social users based on the second feature similarity; deleting the noisy users from the social users; adjusting the second feature of the target user based on the adjusted social relationship information of the target user; and determining the data objects to recommend to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data objects.

13. A method for processing social recommendation models, characterized in that, include: Obtain multiple training samples; The training samples include users' social relationship information, interaction relationship information between users and data objects, users' first features, data objects' first features, and data object recommendation information; The interaction relationship information includes: a set of data objects that interact with the user, and a set of users that interact with the data objects; A network structure for constructing a social recommendation model is provided. This network structure includes a feature encoder, a social graph adjuster, a feature adjuster, and a data object selector. The feature encoder is used to obtain the second features of the user and the data object based on the interaction relationship information, the user's first feature, and the data object's first feature. The social graph adjuster is used to obtain the second feature similarity between users based on the user's second feature and adjust the social relationship information based on the second feature similarity. The feature adjuster is used to adjust the second feature of the target user based on the adjusted social relationship information. The data object selector is used to determine the data object to recommend to the target user based on the similarity between the adjusted second feature of the target user and the second feature of the data object. This includes: obtaining the user's second feature based on a set of data objects that have interacted with the user and the data object's first feature; and obtaining the data object's second feature based on a set of users that have interacted with the data object and the user's first feature. The parameters of the model are learned from the multiple training samples.

14. An electronic device, characterized in that, include: processor; as well as A memory for storing a program for implementing the method according to any one of claims 1-13, wherein the device is powered on and the program for running the method is executed by the processor.

Images