A method, device and system for recommending an interest social relationship, and a storage medium

Abstract

The application provides a kind of interest social relation recommendation method, device, system and storage medium, belong to social recommendation technical field, method includes: import user ID data, original project interaction data and original social data;User ID data and original project interaction data are vectorized to obtain initial user node vector and original project interaction vector;Original project interaction vector and original social data are cleaned to obtain interest social data;Training model is constructed, and training model is analyzed by initial user node vector, original project interaction vector and original interest social data to obtain social relation recommendation model.The application can dynamically filter and remove meaningless social in interest social space, effectively filter social information that can interfere or mislead user, improve the calculation efficiency of recommendation algorithm, also can improve the accuracy of recommendation, has great practical value in recommendation method, greatly meets the needs of user.

Description

Technical Field

[0001] This invention mainly relates to the field of social recommendation technology, specifically to a method, apparatus, system, and storage medium for recommending interest-based social relationships. Background Technology

[0002] Recommendation methods, as a crucial component of today's information age, aim to help users discover and access personalized content that matches their preferences. Collaborative filtering is a common recommendation algorithm that analyzes user behavior and preferences to identify groups of users with similar interests and recommend items or content that might interest them. However, traditional collaborative filtering algorithms perform poorly when facing data sparsity and cold-start problems.

[0003] Most existing recommendation methods based on interest-based social relationships directly utilize raw social data within the social space. However, a significant portion of these relationships is redundant or even noisy. For example, some friends may lack shared interests in specific areas or even have no connection to each other in terms of interests. This renders a large portion of these social relationships meaningless for recommendation tasks, reducing the accuracy and computational efficiency of the recommendation system. Furthermore, existing recommendation methods based on interest-based social relationships, which use vector addition or concatenation, fail to obtain more accurate embeddings of user and item target nodes. Summary of the Invention

[0004] The technical problem to be solved by the present invention is to provide a method, apparatus, system and storage medium for recommending interest-based social relationships, in order to address the shortcomings of the prior art.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: A method for recommending interest-based social relationships, comprising the following steps:

[0006] Import the user ID data corresponding to each target user, the original project interaction data corresponding to each target user, and the original social data corresponding to each target user;

[0007] The user ID data corresponding to each target user and the original project interaction data corresponding to each target user are vectorized to obtain the initial user node vector corresponding to each user ID data and the original project interaction vector corresponding to each original project interaction data.

[0008] Data cleaning is performed on the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user to obtain multiple interest social data corresponding to each target user.

[0009] A training model is constructed, and the training model is analyzed using the initial user node vectors corresponding to all the target users, the multiple original item interaction vectors corresponding to all the target users, and the multiple original interest social data corresponding to all the target users to obtain a social relationship recommendation model.

[0010] Import interaction data from multiple projects to be recommended, and use the social relationship recommendation model to recommend all the interaction data from these projects to obtain the recommendation results for interest-based social relationships.

[0011] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: A recommendation device for interest-based social relationships, comprising:

[0012] The import module is used to import user ID data corresponding to each target user, multiple original project interaction data corresponding to each target user, and multiple original social data corresponding to each target user.

[0013] The vectorization processing module is used to perform vectorization processing on the user ID data corresponding to each target user and the original item interaction data corresponding to each target user, respectively, to obtain the initial user node vector corresponding to each user ID data and the original item interaction vector corresponding to each original item interaction data.

[0014] The data cleaning module is used to clean the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user, so as to obtain multiple interest social data corresponding to each target user.

[0015] The model analysis module is used to build a training model. It analyzes the training model using the initial user node vectors corresponding to all the target users, the multiple original item interaction vectors corresponding to all the target users, and the multiple original interest social data corresponding to all the target users to obtain a social relationship recommendation model.

[0016] The import module is also used to import interaction data from multiple projects to be recommended;

[0017] The recommendation result acquisition module is used to recommend all the interaction data of the items to be recommended through the social relationship recommendation model, and obtain the recommendation result of interest social relationship.

[0018] Based on the above-mentioned method for recommending interest-based social relationships, this invention also provides a recommendation system for interest-based social relationships.

[0019] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a recommendation system for interest-based social relationships, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the interest-based social relationship recommendation method as described above.

[0020] Based on the above-mentioned method for recommending interest-based social relationships, the present invention also provides a computer-readable storage medium.

[0021] Another technical solution of the present invention to solve the above-mentioned technical problems is as follows: a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the interest-based social relationship recommendation method as described above.

[0022] The beneficial effects of this invention are as follows: Initial user node vectors and original item interaction vectors are obtained through vectorization of user ID data and original item interaction data; interest-based social data is obtained through data cleaning of the original item interaction vectors and original social data; a social relationship recommendation model is obtained through model analysis of the trained model using the initial user node vectors, original item interaction vectors, and original interest-based social data; and the recommendation results of interest-based social relationships are obtained through the recommendation of item interaction data to be recommended using the social relationship recommendation model. This invention can dynamically filter and remove meaningless social interactions in the interest-based social space, effectively filtering social information that may interfere with or mislead users, improving the computational efficiency of the recommendation algorithm, and also improving the accuracy of the recommendation. It has significant practical value in recommendation methods and greatly satisfies user needs. Attached Figure Description

[0023] Figure 1 This is one of the flowcharts illustrating a method for recommending interest-based social relationships according to an embodiment of the present invention;

[0024] Figure 2 A second schematic flowchart illustrating a method for recommending interest-based social relationships provided in an embodiment of the present invention;

[0025] Figure 3 This is a block diagram of a recommendation device for interest-based social relationships provided in an embodiment of the present invention. Detailed Implementation

[0026] The principles and features of the present invention are described below with reference to the accompanying drawings. The examples given are only for explaining the present invention and are not intended to limit the scope of the present invention.

[0027] Figure 1 This is a flowchart illustrating a method for recommending interest-based social relationships, as provided in an embodiment of the present invention.

[0028] like Figure 1 As shown, a method for recommending interest-based social relationships includes the following steps:

[0029] Import the user ID data corresponding to each target user, the original project interaction data corresponding to each target user, and the original social data corresponding to each target user;

[0030] The user ID data corresponding to each target user and the original project interaction data corresponding to each target user are vectorized to obtain the initial user node vector corresponding to each user ID data and the original project interaction vector corresponding to each original project interaction data.

[0031] Data cleaning is performed on the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user to obtain multiple interest social data corresponding to each target user.

[0032] A training model is constructed, and the training model is analyzed using the initial user node vectors corresponding to all the target users, the multiple original item interaction vectors corresponding to all the target users, and the multiple original interest social data corresponding to all the target users to obtain a social relationship recommendation model.

[0033] Import interaction data from multiple projects to be recommended, and use the social relationship recommendation model to recommend all the interaction data from these projects to obtain the recommendation results for interest-based social relationships.

[0034] It should be understood that user project interaction data (i.e., raw project interaction data) and raw social data are obtained through the network and then cleaned.

[0035] It should be understood that items are recommended to users based on the trained model (i.e., the social relationship recommendation model).

[0036] Specifically, after obtaining the trained model (i.e., the social relationship recommendation model), the system predicts the user's next item interaction behavior based on the trained model (i.e., the social relationship recommendation model) and recommends items to the user.

[0037] In the above embodiments, initial user node vectors and original project interaction vectors are obtained by vectorizing user ID data and original project interaction data. Data cleaning of the original project interaction vectors and original social data yields interest-based social data. Model analysis of the trained model using the initial user node vectors, original project interaction vectors, and original interest-based social data yields a social relationship recommendation model. The recommendation of interest-based social relationships is then achieved through the recommendation of the project interaction data to be recommended using this social relationship recommendation model. This approach can dynamically filter and remove meaningless social interactions in the interest-based social space, effectively filtering social information that may interfere with or mislead users, improving the computational efficiency of the recommendation algorithm, and also increasing the accuracy of recommendations. It has significant practical value in recommendation methods and greatly satisfies user needs.

[0038] Optionally, as an embodiment of the present invention, the process of cleaning the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user to obtain multiple interest social data corresponding to each target user includes:

[0039] Confidence analysis was performed on the original project interaction vectors corresponding to each target user and the original social data corresponding to each target user to obtain multiple social relationship confidence scores corresponding to each target user.

[0040] The confidence scores of multiple social relationships corresponding to each target user are sorted in descending order to obtain multiple sorted confidence scores of social relationships corresponding to each target user.

[0041] Based on the sorted social relationship confidence scores corresponding to each target user, multiple interest-based social data corresponding to each target user are selected from the multiple original social data corresponding to each target user.

[0042] It should be understood that confidence modeling is performed on all user social relationships (i.e., original project interaction vectors and original social data), and the confidence score (i.e., social relationship confidence score) of each user and all their social relationships is calculated. Each user sorts their original social relationships according to their social confidence score and maps them to the interest social space one by one.

[0043] Specifically, after calculating the confidence score (i.e., social relationship confidence score) for each user and all their social relationships, each user's existing social relationships are sorted according to their social confidence score and mapped one by one to the interest-based social space. In this way, the user's interest relationships in the social network can be presented in the interest-based social space, so as to mine interest-based social relationships and meaningless social interactions in the original social data.

[0044] In the above embodiments, data cleaning is performed on the original project interaction vectors and original social data to obtain interest-based social data. This data can present the user's interest relationships in the social network within the interest-based social space, thereby mining interest-based social relationships and meaningless social interactions in the original social data. This improves the computational efficiency of the recommendation algorithm and also increases the accuracy of the recommendation. It has significant practical value in recommendation methods and greatly satisfies the needs of users.

[0045] Optionally, as an embodiment of the present invention, the original social data includes neighboring users and interaction data of neighboring items corresponding to the neighboring users.

[0046] The process of performing confidence analysis on the original project interaction vectors corresponding to each target user and the original social data corresponding to each target user to obtain confidence scores for multiple social relationships corresponding to each target user includes:

[0047] The interaction data of neighbor items corresponding to each neighbor user are vectorized to obtain the interaction vector of neighbor items corresponding to each neighbor user.

[0048] Each target user is combined with multiple neighboring users to obtain a set of neighboring users corresponding to each target user.

[0049] The confidence scores for multiple social relationships corresponding to each target user are obtained by calculating the confidence scores of each target user through the first formula, which is performed on the original item interaction vectors corresponding to each target user, the set of neighboring users corresponding to each target user, and the multiple neighbor item interaction vectors corresponding to each target user.

[0050]

[0051] Among them, Score iak For target user u i The corresponding interaction vector of the a-th original item and its neighboring user u k The social relationship confidence score between the corresponding neighbor item interaction vectors; Tf() is the Transformer model. For target user u iThe corresponding original project interaction vector of the a-th item. For vector concatenation, For neighboring user u k The neighbor project interaction vector, P u For the target user set, For target user u i The corresponding set of neighboring users.

[0052] Specifically, considering that user-item interaction data to some extent represents user interests, social confidence modeling is performed using historical interaction data between users and their social friends. Furthermore, the Transformer module is well-suited for modeling the similarity between two sequences of user interaction history. Therefore, in the interest-social mapping module, a Transformer-based social confidence modeling method is designed, specifically using the interaction items between users and their social friends to calculate the confidence of their social relationship. The calculation method is shown in the following formula:

[0053]

[0054] in, Indicates user u i The historical interaction item embedding (i.e., the original item interaction vector), Indicates user u j The historical interaction item embedding (i.e., the neighbor item interaction vector), P u This represents the user set (i.e., the target user set). Indicates user u i The original social set (i.e., the set of neighboring users), This represents the concatenation of two vectors, where Tf represents the Transformer module and Score represents the user u. i Its social users u j The social relationship confidence score. The higher the score, the more confident the user is. i Its social users u j The more similar their interests are.

[0055] In the above embodiments, confidence analysis is performed on the original project interaction vectors and original social data to obtain social relationship confidence scores. This can present the user's interest relationships in the social network within the interest social space, improve the computational efficiency of the recommendation algorithm, and also improve the accuracy of the recommendation. It has great practical value in recommendation methods and greatly meets the needs of users.

[0056] Optionally, as an embodiment of the present invention, the process of filtering out multiple interest-based social data corresponding to each target user from multiple original social data corresponding to each target user based on multiple sorted social relationship confidence scores includes:

[0057] The number of interaction data between multiple neighbor items corresponding to each target user is counted to obtain the total number of interactions with neighbor items corresponding to each target user.

[0058] The second formula is used to calculate the total number of interactions with neighboring items corresponding to each target user, thereby obtaining the number of filters corresponding to each target user.

[0059]

[0060] Among them, drop_num i For target user u i The corresponding number of filters, For target user u i The total number of interactions with neighboring items, where ε, α, and β are all hyperparameters;

[0061] The number of social relationship confidence scores corresponding to each target user after sorting is counted to obtain the total number of confidence scores corresponding to each target user;

[0062] The difference between the total confidence score for each target user and the number of filters for each target user is calculated to obtain the number of interest-based social networks for each target user.

[0063] The original social data corresponding to the top A sorted social relationship confidence scores of each target user are used as interest-based social data, thereby obtaining multiple interest-based social data corresponding to each target user, where A is the number of interest-based social data.

[0064] It should be understood that the original number of social connections for each user (i.e., the total number of interactions with neighboring items) is obtained, and then the number of social connections to be removed for each user (drop_num) is adaptively obtained based on this value. Based on the social confidence score (i.e., the confidence score of social relationships after sorting), the low-confidence social relationships in the interest social space are dynamically filtered and removed for each user, thereby obtaining the relatively reliable interest social relationships for each user.

[0065] Specifically, taking the first A original social data points corresponding to each target user as interest-based social data, thus obtaining multiple interest-based social data points corresponding to each target user, can be understood as follows:

[0066] If the first target user 1 has 5 corresponding interest-based social networks and the second target user 2 has 3 corresponding interest-based social networks, then the original social data corresponding to the first 5 sorted social relationship confidence scores of target user 1 will all be used as interest-based social network data. Therefore, the number of interest-based social network data corresponding to target user 1 is 5. Similarly, the original social data corresponding to the first 3 sorted social relationship confidence scores of target user 2 will all be used as interest-based social network data. Therefore, the number of interest-based social network data corresponding to target user 2 is 3.

[0067] Specifically, first, the original number of social connections for each user (i.e., the total number of interactions with neighboring items) is obtained. Then, based on this value, the number of social connections to be removed for each user (drop_num) is adaptively calculated (i.e., the number of filters). Next, based on the social confidence score (i.e., the confidence score of sorted social relationships), duop_num low-confidence social connections are dynamically filtered and removed from each user's interest-based social space, thus removing meaningless social connections. The core idea of ​​this mechanism is that users with sparse social connections suggest retaining all their connections, while users with dense social connections suggest reducing more unreliable connections. This non-uniform discarding method is more robust than uniform discarding that does not consider the number of social connections. The specific method is shown in the following formula:

[0068]

[0069] Where ε, α, and β are three hyperparameters that control the degree of reduction in social relationships, n s The original number of social relationships for a user (i.e., the total number of interactions with neighboring items) is represented by 'drop_num', which represents the number of social relationships that the user suggests removing (i.e., the confidence score of social relationships after sorting). After filtering out meaningless social relationships in the interest-based social space, relatively reliable interest-based social relationships (i.e., interest-based social data) are obtained.

[0070] In the above embodiments, interest-based social data is filtered out from the original social data based on the sorted social relationship confidence scores. This allows for the dynamic filtering and removal of social relationships with low confidence in the interest-based social space, thereby eliminating meaningless social interactions and making interest-based social relationships more reliable, thus improving the accuracy of recommendations.

[0071] Optionally, as an embodiment of the present invention, the process of constructing a training model and performing model analysis on the training model using the initial user node vectors corresponding to all the target users, the multiple original item interaction vectors corresponding to all the target users, and the multiple original interest social data corresponding to all the target users to obtain a social relationship recommendation model includes:

[0072] S41: Construct a training model by training the training model with the initial user node vector corresponding to each target user, multiple original interest social data corresponding to each target user, and multiple original project interaction vectors corresponding to all target users, to obtain multiple predicted scores corresponding to each target user.

[0073] S42: Import multiple real ratings corresponding to each of the target users, calculate the loss value for multiple original item interaction vectors corresponding to all the target users, multiple predicted ratings corresponding to all the target users, and multiple real ratings corresponding to all the target users;

[0074] S43: Update the parameters of the training model according to the loss value to obtain the updated training model, and return to S41 until the preset number of iterations is reached, then use the updated training model as a social relationship recommendation model.

[0075] It should be understood that the model is trained based on the loss function.

[0076] In the above embodiments, a social relationship recommendation model is obtained by performing model analysis on the trained model using the initial user node vector, the original item interaction vector, and the original interest-based social data. This model can dynamically filter and remove meaningless social interactions in the interest-based social space, effectively filtering social information that may interfere with or mislead users, improving the computational efficiency of the recommendation algorithm, and also increasing the accuracy of the recommendation. It has great practical value in recommendation methods and greatly meets the needs of users.

[0077] Optionally, as an embodiment of the present invention, the process of S41 includes:

[0078] The original project interaction vectors corresponding to each target user are respectively aggregated to obtain the original project interaction vector set corresponding to each target user, and the original interest social data corresponding to each target user are respectively aggregated to obtain the original interest social dataset corresponding to each target user. The original project interaction vectors include project embedding sub-vectors and project rating sub-vectors.

[0079] The third equation is used to calculate the original item interaction vector set, the original interest social dataset, the multiple item embedding sub-vectors, and the multiple item rating sub-vectors corresponding to each target user, respectively, to obtain the features of multiple user neighborhood nodes corresponding to each target user. The third equation is:

[0080]

[0081] in, For target user u i The characteristics of the user's neighboring nodes corresponding to the xth user's neighboring nodes. For target user u i The xth user's neighbor node, I(u i ) for target user u i The corresponding original project interaction vector set, S(u i ) for target user u i The corresponding original interest-based social dataset, where W2, W1, b1, and b2 are all learnable training weights, T is the transpose, and W2, b1, b2 ∈ R. d W1∈R d×2d σ() is the ReLU activation function. For vector concatenation, For target user u i The item embedding subvector corresponding to the x-th user's neighborhood node For target user u i The item rating subvector corresponding to the xth user's neighborhood node;

[0082] The fourth equation is used to calculate the initial user node vector corresponding to each target user and the features of multiple user neighbor nodes corresponding to each target user, respectively, to obtain the user neighbor aggregation vector corresponding to each target user. The fourth equation is:

[0083]

[0084] in,

[0085] in,

[0086] in, For target user u i The corresponding user neighborhood aggregation vector, N(u i ) for target user u i The user's neighborhood node set, For target user u i The user importance score corresponding to the xth user's neighboring nodes. For target user u i The characteristics of the user's neighboring nodes corresponding to the xth user's neighboring nodes. For target user u i The user attention score corresponding to the x-th user's neighborhood node, where W6, W5, b5, and b6 are all learnable training weights, and σ() is the ReLU activation function. For vector concatenation, For target user u iThe corresponding initial user node vector,

[0087] The fifth equation is used to calculate the initial user node vector corresponding to each target user and the user neighborhood aggregation vector corresponding to each target user, respectively, to obtain the target user node vector corresponding to each target user. The fifth equation is:

[0088]

[0089] in, For target user u i The corresponding target user node vector, For target user u i The corresponding user neighborhood aggregation vector, For target user u i The corresponding initial user node vectors, sigmoid() is the sigmoid activation function, and tanh() is the tanh activation function. W 10 b 10 W9 and b9 are all trainable parameters. b9, b 10 ∈R d , For vector concatenation, ⊙ represents the Hadamard product;

[0090] Extract multiple original user interaction data corresponding to multiple target projects and project ID data corresponding to each target project from multiple original project interaction data corresponding to all target users;

[0091] The project ID data corresponding to each target project and the original user interaction data corresponding to each target project are vectorized to obtain the initial project node vector corresponding to each project ID data and the original user interaction vector corresponding to each original user interaction data.

[0092] Each target item is combined with multiple original user interaction vectors to obtain an original user interaction vector set for each target item. The original user interaction vectors include user embedding sub-vectors and user rating sub-vectors.

[0093] The sixth equation is used to calculate the original user interaction vector set, multiple user embedding sub-vectors, and multiple user rating sub-vectors corresponding to each target item, respectively, to obtain the features of multiple item neighborhood nodes corresponding to each target item. The sixth equation is:

[0094]

[0095] in, For target project v j The characteristics of the neighboring nodes corresponding to the y-th project's neighboring nodes. For target project v j The y-th neighboring node of the project. U(v j ) for target project v j The corresponding original user interaction vector set, W4, W3, b3, and b4 are all learnable training weights, T is the transpose, and σ() is the ReLU activation function. For vector concatenation, For target project v j The user embedding subvector corresponding to the y-th project's neighborhood node For target project v j The user rating subvector corresponding to the y-th item's neighborhood node;

[0096] The seventh equation is used to calculate the initial project node vector corresponding to each target project and the features of multiple project neighbor nodes corresponding to each target project, thereby obtaining the project neighbor aggregation vector corresponding to each target project. The seventh equation is:

[0097]

[0098] in,

[0099] in,

[0100] in, For target project v j The corresponding project neighborhood aggregation vector, N(v j ) for target project v j Project neighborhood node set For target project v j The project importance score corresponding to the neighboring node of the y-th project. For target project v j The characteristics of the neighboring nodes corresponding to the y-th project's neighboring nodes. For target project v j The project attention score corresponding to the y-th project's neighborhood node, where W8, W7, b7, and b8 are all learnable training weights, and σ() is the ReLU activation function. For vector concatenation, For target project v j The corresponding initial project node vector;

[0101] The target project node vector corresponding to each target project is obtained by calculating the initial project node vector and the project neighborhood aggregation vector corresponding to each target project using the eighth formula. The eighth formula is:

[0102]

[0103] in, For target project v j The corresponding target user node vector, For target project v j The corresponding project neighborhood aggregation vector, For target project v j The corresponding initial item node vectors, sigmoid() is the sigmoid activation function, and tanh() is the tanh activation function. W 12 b 12 W 11 and b 11 All of these are trainable parameters. b 11 b 12 ∈R d , For vector concatenation, ⊙ represents the Hadamard product;

[0104] The ninth formula is used to calculate the target user node vector corresponding to each target user and the target item node vector corresponding to each target item, respectively, to obtain multiple predicted scores corresponding to each target user.

[0105]

[0106] in, For target user u i With target project v j The predicted score, MLP() is a multilayer perceptron, For target user u i The corresponding target user node vector, For target project v j The corresponding target user node vector, This involves concatenating vectors.

[0107] Understandably, a graph neural network is used to aggregate the neighborhood node embeddings (i.e., user neighborhood node features or project neighborhood node features) of user and project target nodes. An attention mechanism is used to aggregate each neighborhood node to obtain the neighborhood aggregated embedding of the target node (i.e., user neighborhood aggregated vector or project neighborhood aggregated vector). Through an adaptive residual attention mechanism, the target node embedding (i.e., initial user node vector or initial project node vector) and the neighborhood aggregated embedding (i.e., user neighborhood aggregated vector or project neighborhood aggregated vector) are flexibly fused to obtain the user embedding representation (i.e., target user node vector) and the project embedding representation (i.e., target project node vector). The prediction layer calculates the user's predicted rating for the project.

[0108] Specifically, a graph neural network is used to aggregate information from the user interaction graph (i.e., the original set of item interaction vectors, item embedding sub-vectors, and item rating sub-vectors) and the interest-based social graph (i.e., the original interest-based social dataset), thereby obtaining the user target node embedding (i.e., the user neighborhood aggregation vector) and the item target node embedding (i.e., the item neighborhood aggregation vector). In the aggregation network, the item target node aggregates item user information and item category information, while the user target node aggregates user interaction information and interest-based social information, and an attention mechanism is used to distinguish the importance of each neighbor node.

[0109] Specifically, to obtain the feature embedding of the target node, the model first needs to obtain the feature embeddings of its neighboring nodes. For the user target node u i The model is based on u i The interaction information (i.e., the original item interaction vector set, item embedding sub-vectors, and item rating sub-vectors) and interest-based social information (i.e., the original interest-based social dataset) are used to extract feature embeddings of neighboring nodes. (i.e., user's neighborhood node characteristics), the specific formula is as follows:

[0110]

[0111] in, Indicates user u i any neighboring node of , and I(u i ) represents user u i The project interaction set, S(u i ) represents user u i Interest-based social networks. Indicates user u i The rating of its neighboring nodes is determined by interaction score when the neighboring node is a project node and interest confidence score when the neighboring node is a user node. and They represent user u respectively iThe embedding vectors of neighboring nodes (i.e., item embedding sub-vectors) and the rating embedding vectors (i.e., item rating sub-vectors). This represents vector concatenation. and W1∈R d×2d These are learnable training weights.

[0112] For project target node v j First, it is necessary to aggregate project type information to obtain the characteristics of the project target node itself, and then use v j The embeddings of neighboring nodes are obtained from the interactive users (i.e., user embedding subvectors) and interactive user ratings (i.e., user rating subvectors). (i.e., the characteristics of the project's neighborhood nodes), the specific formula is as follows:

[0113]

[0114] in, For project v j any neighboring node of , and U(v j ) represents v j The set of interactive users and They represent project v respectively j The embedding vectors of neighboring nodes (i.e., user embedding sub-vectors) and the rating embedding vectors (i.e., user rating sub-vectors).

[0115] It should be understood that after obtaining the neighbor node embeddings of the target node (i.e., user neighbor node features or item neighbor node features), the model still needs to aggregate each neighbor node. For the user target node u i The neighborhood aggregation embedding of the user's target node can be obtained by calculating the following formula. (i.e., the user neighborhood aggregation vector). The specific formula is as follows:

[0116]

[0117]

[0118]

[0119] in, For user target node u i The vector embedding (i.e., the initial user node vector), For user target node u i The neighborhood node embedding (i.e., user neighborhood node features), where σ(·) is the ReLU activation function. This represents the set of neighboring nodes of the user's target node. α represents the attention score. mnUsed to distinguish the importance of each neighboring node. W5, b5, and b6 are learnable training weights.

[0120] For project target node v j The neighborhood aggregation embedding of the project target node can be obtained by calculating the following formula. (i.e., the project neighborhood aggregation vector). The specific formula is as follows:

[0121]

[0122]

[0123]

[0124] in, For the project target node v j The vector embedding (i.e., the initial item node vector), For the project target node v j The neighborhood node embedding (i.e., the feature of the project's neighborhood nodes), where σ(·) is the ReLU activation function. This represents the set of neighboring nodes of the user's target node. α represents the attention score. mn ′ is used to distinguish the importance of each neighboring node.

[0125] Specifically, through the aggregation network, the model has already obtained the neighborhood aggregation features of the target node (i.e., user neighborhood aggregation vector or item neighborhood aggregation vector). It also needs to flexibly fuse the target node features and neighborhood aggregation features through an adaptive residual attention mechanism, so that each user (item) target node obtains a satisfactory final embedding representation (i.e., target user node vector or target item node vector). For user target node u... i The embedding of the user target node can be obtained by calculating the following formula. (i.e., the target user node vector), the specific formula for the adaptive residual attention mechanism is as follows:

[0126]

[0127]

[0128]

[0129] in, Indicates the user's target node u i The vector embedding (i.e., the initial user node vector), Indicates the user's target node u i The neighborhood aggregation embedding (i.e., the user neighborhood aggregation vector), Used to adjust the influence of individual features on the overall features, ⊙ represents the Hadamard product, and tanh(·) and sigmoid(·) represent activation functions. These are trainable parameters.

[0130] For project target node v j The embedding of the project target node can be obtained by calculating the following formula. (i.e., the target project node vector), the specific formula for the adaptive residual attention mechanism is as follows:

[0131]

[0132]

[0133]

[0134] in, Represents the project target node v j The vector embedding (i.e., the initial item node vector), Represents the project target node v j The neighborhood aggregation embedding (i.e., the item neighborhood aggregation vector), Used to adjust the influence of individual features on the overall features, ⊙ represents the Hadamard product, and tanh(·) and sigmoid(·) represent activation functions. These are trainable parameters.

[0135] It should be understood that user u is calculated through the prediction layer. i For project v j Predicted score The formula is as follows:

[0136]

[0137] Among them, MLP(·) is a multilayer perceptron with a 3-layer structure.

[0138] In the above embodiments, the predicted score is obtained by training the training model with the initial user node vector, the original interest social data, and the original item interaction vector. This can dynamically filter and remove meaningless social interactions in the interest social space, effectively filter social information that may interfere with or mislead users, improve the computational efficiency of the recommendation algorithm, and also improve the accuracy of the recommendation. It has great practical value in recommendation methods and greatly meets the needs of users.

[0139] Optionally, as an embodiment of the present invention, in step S42, the process of calculating the loss value for the multiple original item interaction vectors corresponding to all the target users, the multiple predicted ratings corresponding to all the target users, and the multiple real ratings corresponding to all the target users, to obtain the loss value includes:

[0140] The user-item interaction vector set is obtained by combining multiple original item interaction vectors corresponding to all the target users.

[0141] The loss value is obtained by calculating the loss value using the tenth formula on the user item interaction vector set, multiple predicted ratings corresponding to all target users, and multiple real ratings corresponding to all target users. The tenth formula is:

[0142]

[0143] Where Loss is the loss value. For target user u i With target project v j The predicted score, r ij For target user u i With target project v j The actual rating, O is the set of user project interaction vectors, and || is the absolute value.

[0144] Specifically, after obtaining the predicted score, a loss function is needed to train the model. The trained model is obtained when the output loss value stabilizes. The loss function (i.e., the loss value) is defined as follows:

[0145]

[0146] Where |O| is the number of observed user item ratings (i.e., the number of user item interaction vectors in the user item interaction vector set). User u i For project v j The predicted score, r ij User u i For project v j The actual rating.

[0147] In the above embodiments, the loss value is calculated by analyzing the original item interaction vector, predicted score, and actual score, which improves the accuracy of the recommendation and has great practical value in recommendation methods, greatly satisfying the needs of users.

[0148] Optionally, as another embodiment of the present invention, the present invention first constructs a novel interest-based social mapping module, which models the confidence of social relationships based on user interests and maps the original social data to the interest-based social space, thereby gaining a deeper understanding of users' interest relationships in social networks; it constructs a unique social selection mechanism, which uses social confidence scores to dynamically filter and remove meaningless social interactions in the interest-based social space, effectively filtering social information that may interfere with or mislead users; and it uses an augmented graph neural network to model a heterogeneous relationship graph composed of interest-based social relationships and user-item relationships, obtaining user target node embeddings and item target node embeddings, and then feeding both into the prediction layer for final score prediction.

[0149] Alternatively, as another embodiment of the present invention, deep neural network technology for graph data has made significant progress. Graph neural networks can effectively integrate node information and topological structure, possessing powerful capabilities in learning graph structure data representations. Since user interaction data and user social data can both be represented as graph structure data in social recommendation, graph neural networks have been widely applied in social recommendation and rating prediction tasks, providing new avenues for better learning feature representations of users and items.

[0150] Optionally, as another embodiment of the present invention, in order to reliably mine meaningless social relationships in raw social data, the present invention first constructs an interest-based social mapping module. This module uses Transformer to model the confidence level of social relationships based on user interests and maps the raw social data to an interest-based social space, thereby gaining a deeper understanding of users' interest relationships in social networks. Then, combined with a social selection mechanism, meaningless social relationships in the interest-based social space are dynamically filtered and removed using social confidence scores. This effectively filters out social information that may interfere with or mislead users, retaining only meaningful interest-based social relationships.

[0151] To obtain more accurate user and project target node embeddings, we first use an enhanced graph neural network for modeling. Project target nodes aggregate project user information and project category information, while user target nodes aggregate user interaction information and interest-based social information. An attention mechanism is used to distinguish the importance of each neighbor node. Then, an adaptive residual attention mechanism is used to fuse the features to obtain more accurate user target node embeddings and project target node embeddings, thereby improving the accuracy of the recommendation system.

[0152] Alternatively, as another embodiment of the present invention, the beneficial effects of the present invention are as follows:

[0153] This technology can effectively map raw social data from the social space to an interest-based social space, retaining only meaningful interest-based social relationships and generating a more concise interest-based social network. This feature not only helps improve the computational efficiency of recommendation algorithms but also increases recommendation accuracy, making it highly valuable in recommendation systems.

[0154] Alternatively, as another embodiment of the present invention, such as Figure 2 As shown, the specific steps of the present invention are as follows:

[0155] Step S1: Obtain user project interaction data and raw social data through the network, and perform data cleaning;

[0156] Step S2: Use the social mapping module to model the confidence of all users' social relationships and calculate the confidence score of each user and all their social relationships;

[0157] Step S3: Each user sorts their existing social relationships according to their social confidence score and maps them to interest-based social spaces one by one;

[0158] Step S4: Obtain the original number of social contacts for each user, and then adaptively obtain the number of social contacts to be removed for each user, drop_num, based on this value;

[0159] Step S5: Based on the social confidence score, dynamically filter and remove duop_num low-confidence social relationships in the interest-based social space for each user, thereby obtaining relatively reliable interest-based social relationships for each user. Use a graph neural network to aggregate information from the user interaction graph and the interest-based social graph to obtain the user target node embedding and the project target node embedding.

[0160] Step S6: Use a graph neural network to aggregate the neighbor node embeddings of user and project target nodes;

[0161] Step S7: Aggregate all neighboring nodes using an attention mechanism to obtain the neighborhood aggregation embedding of the target node;

[0162] Step S8: Through an adaptive residual attention mechanism, the target node embedding and the neighborhood aggregation embedding are flexibly fused to obtain the user embedding representation and the item embedding representation;

[0163] Step S9: The prediction layer calculates the user's predicted rating for the project;

[0164] Step S10: Train the model based on the loss function;

[0165] Step S11: Recommend items to the user based on the trained model.

[0166] Figure 3This is a block diagram of a recommendation device for interest-based social relationships provided in an embodiment of the present invention.

[0167] Alternatively, as another embodiment of the present invention, such as Figure 3 As shown, a recommendation device for interest-based social relationships includes:

[0168] The import module is used to import user ID data corresponding to each target user, multiple original project interaction data corresponding to each target user, and multiple original social data corresponding to each target user.

[0169] The vectorization processing module is used to perform vectorization processing on the user ID data corresponding to each target user and the original item interaction data corresponding to each target user, respectively, to obtain the initial user node vector corresponding to each user ID data and the original item interaction vector corresponding to each original item interaction data.

[0170] The data cleaning module is used to clean the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user, so as to obtain multiple interest social data corresponding to each target user.

[0171] The model analysis module is used to build a training model. It analyzes the training model using the initial user node vectors corresponding to all the target users, the multiple original item interaction vectors corresponding to all the target users, and the multiple original interest social data corresponding to all the target users to obtain a social relationship recommendation model.

[0172] The import module is also used to import interaction data from multiple projects to be recommended;

[0173] The recommendation result acquisition module is used to recommend all the interaction data of the items to be recommended through the social relationship recommendation model, and obtain the recommendation result of interest social relationship.

[0174] Optionally, another embodiment of the present invention provides a recommendation system for interest-based social relationships, including a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the interest-based social relationship recommendation method as described above. This system can be a computer or similar system.

[0175] Optionally, another embodiment of the present invention provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the method for recommending interest-based social relationships as described above.

[0176] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0177] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working process of the above-described apparatus and unit can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0178] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative. For instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0179] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments of the present invention, depending on actual needs.

[0180] Furthermore, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0181] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0182] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for recommending interest-based social relationships, characterized in that, Includes the following steps: Import the user ID data corresponding to each target user, the original project interaction data corresponding to each target user, and the original social data corresponding to each target user; The user ID data corresponding to each target user and the original project interaction data corresponding to each target user are vectorized to obtain the initial user node vector corresponding to each user ID data and the original project interaction vector corresponding to each original project interaction data. Data cleaning is performed on the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user to obtain multiple interest social data corresponding to each target user. A training model is constructed, and the training model is analyzed using the initial user node vectors corresponding to all the target users, the multiple original item interaction vectors corresponding to all the target users, and the multiple original interest social data corresponding to all the target users to obtain a social relationship recommendation model. Import interaction data from multiple projects to be recommended, and use the social relationship recommendation model to recommend all the interaction data from these projects to obtain the recommendation results for interest-based social relationships.

2. The method for recommending interest-based social relationships according to claim 1, characterized in that, The process of cleaning the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user to obtain multiple interest-based social data corresponding to each target user includes: Confidence analysis was performed on the original project interaction vectors corresponding to each target user and the original social data corresponding to each target user to obtain multiple social relationship confidence scores corresponding to each target user. The confidence scores of multiple social relationships corresponding to each target user are sorted in descending order to obtain multiple sorted confidence scores of social relationships corresponding to each target user. Based on the sorted social relationship confidence scores corresponding to each target user, multiple interest-based social data corresponding to each target user are selected from the multiple original social data corresponding to each target user.

3. The method for recommending interest-based social relationships according to claim 2, characterized in that, The raw social data includes neighboring users and interaction data of neighboring items corresponding to those neighboring users. The process of performing confidence analysis on the original project interaction vectors corresponding to each target user and the original social data corresponding to each target user to obtain multiple social relationship confidence scores corresponding to each target user includes: The interaction data of neighbor items corresponding to each neighbor user are vectorized to obtain the interaction vector of neighbor items corresponding to each neighbor user. Each target user is combined with multiple neighboring users to obtain a set of neighboring users corresponding to each target user. The confidence scores for multiple social relationships corresponding to each target user are obtained by calculating the confidence scores of each target user through the first formula, which is performed on the original item interaction vectors corresponding to each target user, the set of neighboring users corresponding to each target user, and the multiple neighbor item interaction vectors corresponding to each target user. Among them, Score iak For target user u i The corresponding interaction vector of the a-th original item and its neighboring user u k The social relationship confidence score between the corresponding neighbor item interaction vectors; Tf() is the Transformer model. For target user u i The corresponding original project interaction vector of the a-th item. For vector concatenation, For neighboring user u k The neighbor project interaction vector, P u For the target user set, For target user u i The corresponding set of neighboring users.

4. The method for recommending interest-based social relationships according to claim 3, characterized in that, The process of filtering out multiple interest-based social data corresponding to each target user from multiple original social data corresponding to each target user based on multiple sorted social relationship confidence scores includes: The number of interaction data between multiple neighbor items corresponding to each target user is counted to obtain the total number of interactions with neighbor items corresponding to each target user. The second formula is used to calculate the total number of interactions with neighboring items corresponding to each target user, thereby obtaining the number of filters corresponding to each target user. Among them, drop_num i For target user u i The corresponding number of filters, For target user u i The total number of interactions with neighboring items, where ε, α, and β are all hyperparameters; The number of social relationship confidence scores corresponding to each target user after sorting is counted to obtain the total number of confidence scores corresponding to each target user; The difference between the total confidence score for each target user and the number of filters for each target user is calculated to obtain the number of interest-based social networks for each target user. The original social data corresponding to the top A sorted social relationship confidence scores of each target user are used as interest-based social data, thereby obtaining multiple interest-based social data corresponding to each target user, where A is the number of interest-based social data.

5. The method for recommending interest-based social relationships according to claim 1, characterized in that, The process of constructing and training a model, which involves analyzing the training model using initial user node vectors corresponding to all target users, multiple original item interaction vectors corresponding to all target users, and multiple original interest-based social data corresponding to all target users, to obtain a social relationship recommendation model, includes: S41: Construct a training model by training the training model with the initial user node vector corresponding to each target user, multiple original interest social data corresponding to each target user, and multiple original project interaction vectors corresponding to all target users, to obtain multiple predicted scores corresponding to each target user. S42: Import multiple real ratings corresponding to each of the target users, calculate the loss value for multiple original item interaction vectors corresponding to all the target users, multiple predicted ratings corresponding to all the target users, and multiple real ratings corresponding to all the target users; S43: Update the parameters of the training model according to the loss value to obtain the updated training model, and return to S41 until the preset number of iterations is reached, then use the updated training model as a social relationship recommendation model.

6. The method for recommending interest-based social relationships according to claim 5, characterized in that, The process in S41 includes: The original project interaction vectors corresponding to each target user are respectively aggregated to obtain the original project interaction vector set corresponding to each target user, and the original interest social data corresponding to each target user are respectively aggregated to obtain the original interest social dataset corresponding to each target user. The original project interaction vectors include project embedding sub-vectors and project rating sub-vectors. The third equation is used to calculate the original item interaction vector set, the original interest social dataset, the multiple item embedding sub-vectors, and the multiple item rating sub-vectors corresponding to each target user, respectively, to obtain the features of multiple user neighborhood nodes corresponding to each target user. The third equation is: in, For target user u i The characteristics of the user's neighboring nodes corresponding to the xth user's neighboring nodes. For target user u i The xth user's neighbor node, I(u i ) for target user u i The corresponding original project interaction vector set, S(u i ) for target user u i The corresponding original interest-based social dataset, where W2, W1, b1, and b2 are all learnable training weights, T is the transpose, and W2, b1, b2 ∈ R. d W1∈R d×2d σ() is the ReLU activation function. For vector concatenation, For target user u i The item embedding subvector corresponding to the x-th user's neighborhood node For target user u i The item rating subvector corresponding to the xth user's neighborhood node; The fourth equation is used to calculate the initial user node vector corresponding to each target user and the features of multiple user neighbor nodes corresponding to each target user, respectively, to obtain the user neighbor aggregation vector corresponding to each target user. The fourth equation is: in, in, in, For target user u i The corresponding user neighborhood aggregation vector, N(u i ) for target user u i The user's neighborhood node set, For target user u i The user importance score corresponding to the xth user's neighboring nodes. For target user u i The characteristics of the user's neighboring nodes corresponding to the xth user's neighboring nodes. For target user u i The user attention score corresponding to the x-th user's neighborhood node, where W6, W5, b5, and b6 are all learnable training weights, and σ() is the ReLU activation function. For vector concatenation, For target user u i The corresponding initial user node vector, The fifth equation is used to calculate the initial user node vector corresponding to each target user and the user neighborhood aggregation vector corresponding to each target user, respectively, to obtain the target user node vector corresponding to each target user. The fifth equation is: in, For target user u i The corresponding target user node vector, For target user u i The corresponding user neighborhood aggregation vector, For target user u i The corresponding initial user node vectors, sigmoid() is the sigmoid activation function, and tanh() is the tanh activation function. W 10 b 10 W9 and b9 are all trainable parameters. For vector concatenation, ⊙ represents the Hadamard product; Extract multiple original user interaction data corresponding to multiple target projects and project ID data corresponding to each target project from multiple original project interaction data corresponding to all target users; The project ID data corresponding to each target project and the original user interaction data corresponding to each target project are vectorized to obtain the initial project node vector corresponding to each project ID data and the original user interaction vector corresponding to each original user interaction data. Each target item is combined with multiple original user interaction vectors to obtain an original user interaction vector set for each target item. The original user interaction vectors include user embedding sub-vectors and user rating sub-vectors. The sixth equation is used to calculate the original user interaction vector set, multiple user embedding sub-vectors, and multiple user rating sub-vectors corresponding to each target item, respectively, to obtain the features of multiple item neighborhood nodes corresponding to each target item. The sixth equation is: in, For target project v j The characteristics of the neighboring nodes corresponding to the y-th project's neighboring nodes. For target project v j The y-th neighboring node of the project. U(v j ) for target project v j The corresponding original user interaction vector set, W4, W3, b3, and b4 are all learnable training weights, T is the transpose, and σ() is the ReLU activation function. For vector concatenation, For target project v j The user embedding subvector corresponding to the y-th project's neighborhood node For target project v j The user rating subvector corresponding to the y-th item's neighborhood node; The seventh equation is used to calculate the initial project node vector corresponding to each target project and the features of multiple project neighbor nodes corresponding to each target project, thereby obtaining the project neighbor aggregation vector corresponding to each target project. The seventh equation is: in, in, in, For target project v j The corresponding project neighborhood aggregation vector, N(v j ) for target project v j Project neighborhood node set For target project v j The project importance score corresponding to the neighboring node of the y-th project. For target project v j The characteristics of the neighboring nodes corresponding to the y-th project's neighboring nodes. For target project v j The project attention score corresponding to the y-th project's neighborhood node, where W8, W7, b7, and b8 are all learnable training weights, and σ() is the ReLU activation function. For vector concatenation, For target project v j The corresponding initial project node vector; The target project node vector corresponding to each target project is obtained by calculating the initial project node vector and the project neighborhood aggregation vector corresponding to each target project using the eighth formula. The eighth formula is: in, For target project v j The corresponding target user node vector, For target project v j The corresponding project neighborhood aggregation vector, For target project v j The corresponding initial item node vectors, sigmoid() is the sigmoid activation function, and tanh() is the tanh activation function. W 12 b 12 W 11 and b 11 All of these are trainable parameters. For vector concatenation, ⊙ represents the Hadamard product; The ninth formula is used to calculate the target user node vector corresponding to each target user and the target item node vector corresponding to each target item, respectively, to obtain multiple predicted scores corresponding to each target user. in, For target user u i With target project v j The predicted score, MLP() is a multilayer perceptron, For target user u i The corresponding target user node vector, For target project v j The corresponding target user node vector, This involves concatenating vectors.

7. The method for recommending interest-based social relationships according to claim 6, characterized in that, In step S42, the process of calculating the loss value for multiple original item interaction vectors corresponding to all target users, multiple predicted ratings corresponding to all target users, and multiple real ratings corresponding to all target users includes: The user-item interaction vector set is obtained by combining multiple original item interaction vectors corresponding to all the target users. The loss value is obtained by calculating the loss value using the tenth formula on the user item interaction vector set, multiple predicted ratings corresponding to all target users, and multiple real ratings corresponding to all target users. The tenth formula is: Where Loss is the loss value. For target user u i With target project v j The predicted score, r ij For target user u i With target project v j The actual rating, O is the set of user project interaction vectors, and || is the absolute value.

8. A recommendation device for interest-based social relationships, characterized in that, include: The import module is used to import user ID data corresponding to each target user, multiple original project interaction data corresponding to each target user, and multiple original social data corresponding to each target user. The vectorization processing module is used to perform vectorization processing on the user ID data corresponding to each target user and the original item interaction data corresponding to each target user, respectively, to obtain the initial user node vector corresponding to each user ID data and the original item interaction vector corresponding to each original item interaction data. The data cleaning module is used to clean the multiple original project interaction vectors corresponding to each target user and the multiple original social data corresponding to each target user, so as to obtain multiple interest social data corresponding to each target user. The model analysis module is used to build a training model. It analyzes the training model using the initial user node vectors corresponding to all the target users, the multiple original item interaction vectors corresponding to all the target users, and the multiple original interest social data corresponding to all the target users to obtain a social relationship recommendation model. The import module is also used to import interaction data from multiple projects to be recommended; The recommendation result acquisition module is used to recommend all the interaction data of the items to be recommended through the social relationship recommendation model, and obtain the recommendation result of interest social relationship.

9. A recommendation system for interest-based social relationships, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method for recommending interest-based social relationships as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by a processor, it implements the method for recommending interest-based social relationships as described in any one of claims 1 to 7.

Images