A performance optimization method for social recommendation system based on social importance and item category information

Abstract

The present application relates to a method for optimizing the performance of a social recommendation system based on social importance and project category information. The method is implemented based on a graph neural network. Based on the user social network graph SG, the user-project graph UIG and the project-category graph ICG, a similar user model, a social importance model and a project category information model for cold-start users are established. The user-project global information is fused through the above models. After fusion, the prediction model is used to predict and score the uninteracted projects, and finally the project recommendation is completed for the user. The present application significantly enhances the preference embedding representation of ordinary users and cold-start users, enriches the semantic representation of users, and provides richer semantic connotations for users and projects.

Description

Technical Field

[0001] The present application relates to the technical field of social system modeling, and in particular to a method for optimizing the performance of a social recommendation system based on social importance and project category information. Background Art

[0002] How online platforms filter a large amount of redundant information to provide personalized services to platform users is the core of their market competitiveness. Enhancing the performance of recommendation systems is an effective means for platforms to provide personalized services, and making full use of social relationships is an effective measure to enhance the performance of recommendation systems. In practice, recommendation systems use graph neural networks with powerful representation learning capabilities, which can simultaneously aggregate social network and user-item interaction network information, achieve good representation for user and item learning, and improve recommendation performance.

[0003] Although the recommendation system based on graph neural network has obvious advantages, it still has the following problems: (1) It lacks attention to users with less project interaction information and social information. (2) The user's rating weight of the project does not take into account its importance in the social network. When the project information is spread through the social network, the rating weight of the user who interacts with it at the hub node of the social network is significantly greater than the rating weight of the user who interacts with it but is at the edge of the social network. (3) The mining of user-project information is insufficient, and it is difficult to improve the performance of the recommendation system through other information other than social network information. Summary of the invention

[0004] The present application provides a method for optimizing the performance of a social recommendation system based on social importance and project category information, which solves the problems of a recommendation system based on graph neural networks, such as a lack of mining of effective preference information of cold-start users, failure to differentiate between user rating weights for projects, and insufficient mining of user project category information.

[0005] A method for optimizing the performance of a social recommendation system based on social importance and item category information. The method is implemented based on a social network graph, a user-item graph, and an item-category graph. The social network graph is used to describe the social relationship between users, the user-item graph is used to describe the interactive relationship between users and items, and the item-category graph is used to describe the category to which the item belongs. The method specifically includes the following steps:

[0006] S01) Establish similar user models for cold start users;

[0007] The project purchase record of each user in the social network graph is obtained, and the cold start user is defined based on the number of interactions between the user and its own node and project; users with the same purchase preference as the cold start user are obtained based on the historical interaction data of the user and the project, and the users with the same purchase preference are used as similar users of the cold start user; based on the social network graph of the cold start user and the similar users of the cold start user, a social network enhanced graph is generated; in the social network enhanced graph, the first-order neighbor users of the user constitute the social space of the user, the first-order projects of the user constitute the project space of the user, and the first-order users of the project constitute the user space of the project;

[0008] S02) Build a social importance model;

[0009] The PageRank value of each user in the social network is P The integer part of the product of is taken as the user evaluation weight, which is used to characterize the importance of the user in the social network; P is the smallest positive integer that makes the product greater than 1; for users without social information, their user evaluation weight is 1;

[0010] Latent representation of users in social importance model , including the latent representation of users in the item space in building social importance models and the latent representation of users in the social space for building social importance models ; The formula for obtaining is:

[0011] ;

[0012] In the formula, represents a multi-layer perceptron, Represents a vector concatenation operation;

[0013] in, The latent representation of the item space in the social importance model is acquired through graph neural network learning. Acquire the latent representation of social space in social importance model through graph neural network learning;

[0014] Latent representation of items in social importance model , the latent representation acquisition of user space in social importance model is established through graph neural network learning;

[0015] S03) Establish project category information model;

[0016] User potential representation of item category information model , including the potential representation of users in the item space in building the item category information model and users’ latent representation in social space for modeling item category information ; The formula for obtaining is:

[0017] ;

[0018] in, The potential representation of the project space in the project category information model is acquired through graph neural network learning. Acquire the latent representation of social space in the project category information model through graph neural network learning;

[0019] Project Potential Representation of Project Category Information Model Acquire the potential representation of user space in the project category information model through graph neural network learning;

[0020] S04) User-item global information fusion;

[0021] The final embedding representation of the user is obtained by fusing the social importance model and the user potential representation in the item category information through a multi-layer perceptron. ;

[0022] The final embedding representation of the project is obtained by fusing the social importance model and the potential representation of the project in the project category information through a multi-layer perceptron ;

[0023] S05 predicts the user's ratings for uninteracted items;

[0024] Splicing and , input the concatenated representation into the multi-layer perceptron to obtain the user About Project The predicted score value of ;

[0025] The user's rating of the non-interacted items is obtained by training the prediction model, and the objective function of the prediction model is:

[0026] ;

[0027] In the formula, Represents the training sample set the number of Indicates user About Project The predicted score value of Indicates user About Project The actual rating value.

[0028] Furthermore, in step S01), a cold start user is defined as a user whose number of interactions with its neighboring nodes and projects is less than K The next node, K Is a positive integer.

[0029] Further, in step S02), The acquisition method is:

[0030] Learning latent representations of items in item space via graph neural networks , aggregate using aggregate functions , and introduce an attention mechanism in the aggregation function to obtain the weights of different items for user feature representation, where the user feature is a multi-dimensional representation of user embedding; the specific formula is:

[0031] ;

[0032] In the formula, Represents the Relu activation function; Represent the weights and biases of the neural network respectively; Indicates user Interactive Project Latent representation in item space, The dimension is d A real vector of ; Represents an aggregate function, which is an addition function; Indicates user Interactive Project Collection the number of

[0033] in, ;

[0034] ;

[0035] In the formula, Indicates user Interactive Project The initial embedding representation of Indicates user The initial embedding representation of user rating weights, Indicates user Interactive Project The initial embedding representation of the evaluation; Indicates user The initial embedding representation obtained based on the status and ratings in the social network, Represents the Tensor product symbol of two vectors;

[0036] After introducing the attention mechanism:

[0037] ;

[0038] In the formula, represents the weights of different interaction items for user feature representation after normalization using the Softmax function; and:

[0039] ;

[0040] ;

[0041] in, Indicates user The initial embedding representation of represents the weight of the neural network, Represents the bias of the neural network.

[0042] Further, in step S02), The method to obtain is:

[0043] Learning latent representations of first-order neighbor users in item space via graph neural networks ,by As the first-order neighbor user-item purchase preference embedding representation, use the aggregation function to aggregate , and introduce the attention mechanism into the aggregation function to obtain the weights of different first-order neighbor users for user feature representation; the specific formula is:

[0044] ;

[0045] In the formula, Indicates user The first-order neighbor nodes Item purchase preference embedding representation;

[0046] After introducing the attention mechanism:

[0047] ；

[0048] In the formula, represents the weights of different first-order neighbor users for user feature representation after normalization using the Softmax function, and:

[0049] ;

[0050] .

[0051] Further, in step S02), The method to obtain is:

[0052] Learning latent representations of users interacting with items in user space via graph neural networks , aggregate using aggregate functions , and introduce the attention mechanism in the aggregation function to obtain the weights of different users' potential representations of items; the specific formula is:

[0053] ;

[0054] In the formula, Display items A latent representation of the interacting user in user space; Display items Interactive user collection the number of Represents an aggregate function, which is an addition function;

[0055] in, The method to obtain is:

[0056] ;

[0057] ;

[0058] In the formula, Indicates user The initial embedding representation of Indicates user In the initial embedding representation of user rating weights, Indicates user About Project The initial embedding representation of the evaluation;

[0059] After introducing the attention mechanism:

[0060] ;

[0061] In the formula, represents the weights of different users' potential representations of items after normalization using the Softmax function, and:

[0062] ;

[0063] .

[0064] Further, in step S03), The formula for obtaining is:

[0065] ;

[0066] In the formula, Indicates user Interactive Project Latent representation in item space;

[0067] in, The formula for obtaining is:

[0068] ;

[0069] ;

[0070] ;

[0071] In the formula, Indicates user Interactive Project The initial embedding representation of the category; Indicates user About Project The initial embedding representation of the rating is Indicates user About Project Initial embedding representation of category preferences;

[0072] Introduce the attention mechanism in the aggregation function:

[0073] ;

[0074] in, represents the weights of different interaction items for user feature representation after normalization using the Softmax function, and:

[0075] ;

[0076] .

[0077] Further, in step S03), The formula for obtaining is:

[0078] ;

[0079] in, Indicates user The first-order neighbor nodes Item purchase preference embedding;

[0080] After introducing the attention mechanism:

[0081] ;

[0082] In the formula, represents the weights of different first-order neighbor users for user feature representation after normalization using the Softmax function, and:

[0083] ;

[0084] .

[0085] Further, in step S03), The formula for obtaining is:

[0086] ;

[0087] In the formula, Display items A latent representation of the interacting user in user space;

[0088] After introducing the attention mechanism:

[0089] ;

[0090] In the formula, represents the weights of the potential representations of different users to items after normalization using the Softmax function, and:

[0091] ;

[0092] .

[0093] The beneficial effects of this application are:

[0094] 1. In step S01), the present application adds similar users to the cold-start user based on the historical interaction data between the user and the project. The historical interaction data is based on the project purchase preference, which enhances the social information embedding representation of the cold-start user.

[0095] 2. When project information is propagated through social networks, the rating weight of users who interact with it at the hub nodes of the social network is significantly greater than the rating weight of marginal users. In step S02), the present application adds user evaluation weights, takes into account the centrality of users in the social network, and enriches the semantic representation of users.

[0096] 3. In the process of modeling the project category information model, the application adds the category information of the project interacted by the user. The introduction of the project category information can provide richer semantic connotations for users and projects.

[0097] 4. In this application, the social importance model and the project category information model have similarities and differences. The similarities are: the user embedding potential representations all come from the user's project space and the user's social space; the project potential representations all come from the user-project interaction network. The difference is: the model emphasizes different attribute cores. The former emphasizes the differences in the influence of users on the dissemination of project evaluations in social networks, while the latter highlights the user preference information rich in project category information and the similarities between projects of the same category and the differences between projects of different categories; in addition, the embedding and propagation methods of the two are also different. BRIEF DESCRIPTION OF THE DRAWINGS

[0098] Figure 1 Enhance the graph for the user social network in this application;

[0099] Figure 2 A global interaction category graph formed by the social network enhancement graph, user-item graph, and item-category graph in this application;

[0100] Figure 3 Schematic diagram of the learning process of user and item latent representation based on social importance modeling;

[0101] Figure 4 Schematic diagram of the user and item latent representation learning process based on item category information modeling. DETAILED DESCRIPTION

[0102] Based on the problems existing in the recommendation system of the graph neural network described in the background technology, this application provides a performance optimization method for a social recommendation system based on social importance and item category information. This method is implemented based on a social network graph (SG), a user-item graph (UIG), and an item-category graph (ICG). The social network graph is used to describe the social relationship between users, the user-item graph is used to describe the interactive relationship between users and items, and the item-category graph is used to describe the category to which the item belongs. The users in the SG are consistent with the users in the UIG, and the items in the UIG are consistent with the items in the ICG.

[0103] In this embodiment, , as well as Represent the user set, item set and item category set respectively, where N is the number of users, M is the number of items, and H is the number of categories. Then SG, UIG, and ICG can be represented by sets: , , .

[0104] Specifically, social recommendation is different from traditional task recommendation based on user-item interaction data. Social recommendation regards social relationships between users as auxiliary information to help learn user preferences. In social recommendation, social relationships come from social networks composed of users with follower relationships (follow) or friend relationships. With the assistance of SG, the goal of social recommendation tasks is the same as that of traditional tasks. This application establishes a prediction model, predicts the user's score for the project through the prediction model, and then selects the top-ranked project as the recommendation result for the user.

[0105] This method specifically comprises the following steps:

[0106] S01) Establish similar user models for cold start users;

[0107] The project purchase record of each user in the social network graph is obtained, and the cold-start user is defined based on the number of interactions between the user and its own node and project; users with the same purchase preference as the cold-start user are obtained based on the historical interaction data of the user and the project, and the users with the same purchase preference are regarded as similar users of the cold-start user; based on the social network graph of the cold-start user and the similar users of the cold-start user, an enhanced social network graph is generated; in the enhanced social network graph, the first-order neighbor users of the user constitute the social space of the user, the first-order projects of the user constitute the project space of the user, and the first-order users of the project constitute the user space of the project.

[0108] As attached Figure 1 As shown, the cold start user in this application is established based on the user's interaction data. It is a dynamic feature that defines the cold start user from the user's purchase preference consistency level. In this embodiment, the cold start user is defined as a user whose number of interactions with its adjacent nodes and items is less than K The next node, K In this embodiment, the value is 10. The present application is not limited to the definition of cold start users, but highlights the influence of the user's social information on the accuracy of recommendation by defining cold start users.

[0109] In this application, the process of similar user modeling is as follows: First, obtain the purchase records of each user in the social network to form a similar user model with a length of s Purchase vector , s Represents the number of item categories; then all users' purchase vectors Composition Matrix F , F Each row in represents a ; Next, in the social network graph, nodes whose number of interactions with their adjacent nodes and projects is less than 10 are obtained as cold start users; for each cold start user, calculate and The tensor product of X users as similar users of cold start users ( X is a custom parameter).

[0110] In this application, SUG is a graph generated by modeling similar users. The nodes in SUG are a subset of the nodes in SG. The edges are directed from users to their similar users. The edges in SUG are added to SG to obtain the social network enhanced graph SNEG.

[0111] S02) Build a social importance model.

[0112] In the social importance model, the representation of a user in the model comes from the fusion of the user-item space representation and the user's social space representation, and the representation of an item in the model comes from learning from interactions with users.

[0113] The PageRank value of each user in the social network is P The integer part of the product of is taken as the user evaluation weight, which is used to characterize the importance of the user in the social network; P The smallest positive integer that makes the product greater than 1; for users without social information, their user evaluation weight is 1. The PageRank value is multiplied by 10 P The purpose is only to make the value of the user weight greater than 1. This application applies the existing PageRank algorithm to the Internet network. The Internet network is a directed network. The more nodes are pointed to, the higher their importance in the network.

[0114] This application assigns different evaluation weights to users from the perspective of social network hubness. User potential representation of social importance model Including users’ latent representation in item space in building social importance models and the latent representation of users in the social space for building social importance models ; The formula for obtaining is:

[0115] ;

[0116] In the formula, represents a multi-layer perceptron, Represents a vector concatenation operation.

[0117] in, The latent representation of the item space in the social importance model is acquired through graph neural network learning. Acquiring latent representations of social space in social importance modeling via graph neural network learning.

[0118] The acquisition method is:

[0119] Learning latent representations of items in item space via graph neural networks , aggregate using aggregate functions , and introduce an attention mechanism in the aggregation function to obtain the weights of different items for user feature representation, where the user feature is a multi-dimensional representation of user embedding; the specific formula is:

[0120] ;

[0121] In the formula, Represents the Relu activation function; Represent the weights and biases of the neural network respectively; Indicates user Interactive Project Latent representation in item space, The dimension is d A real vector of ; Represents an aggregate function, which is an addition function; Indicates user Interactive Project Collection the number of

[0122] in, ;

[0123] ;

[0124] In the formula, Indicates user Interactive Project The initial embedding representation of Indicates user The initial embedding representation of user rating weights, Indicates user Interactive Project The initial embedding representation of the evaluation; Indicates user The initial embedding representation obtained based on the status and ratings in the social network, The symbol for the product of two vectors;

[0125] After introducing the attention mechanism:

[0126] ;

[0127] In the formula, represents the weights of different interaction items for user feature representation after normalization using the Softmax function; and:

[0128] ;

[0129] ;

[0130] in, Indicates user The initial embedding representation of represents the weight of the neural network, Represents the bias of the neural network.

[0131] Since the user's behavior will be affected by the preferences of neighbor nodes in the social network, this application uses the latent representation of the user's neighbor nodes in the project space as their preference embedding. The method to obtain is:

[0132] Learning latent representations of first-order neighbor users in item space via graph neural networks ,by As the first-order neighbor user-item purchase preference embedding representation, use the aggregation function to aggregate , and introduce the attention mechanism into the aggregation function to obtain the weights of different first-order neighbor users for user feature representation; the specific formula is:

[0133] ;

[0134] In the formula, Indicates user The first-order neighbor nodes Item purchase preference embedding representation;

[0135] After introducing the attention mechanism:

[0136] ；

[0137] In the formula, represents the weights of different first-order neighbor users for user feature representation after normalization using the Softmax function, and:

[0138] ;

[0139] .

[0140] Item Latent Representation for Social Importance Model Learning latent representations in user space via application networks. The method to obtain is:

[0141] Learning latent representations of users interacting with items in user space via graph neural networks , aggregate using aggregate functions , and introduce the attention mechanism in the aggregation function to obtain the weights of different users' potential representations of items; the specific formula is:

[0142] ;

[0143] In the formula, Display items A latent representation of the interacting user in user space; Display items Interactive user collection the number of Represents an aggregate function, which is an addition function;

[0144] in, The method to obtain is:

[0145] ;

[0146] ;

[0147] In the formula, Indicates user The initial embedding representation of Indicates user In the initial embedding representation of user rating weights, Indicates user About Project The initial embedding representation of the evaluation;

[0148] After introducing the attention mechanism:

[0149] ;

[0150] In the formula, represents the weights of different users' potential representations of items after normalization using the Softmax function, and:

[0151] ;

[0152] .

[0153] S03) Establish a project category information model. In order to enrich the interactive information between users and projects, so as to better capture the user's preference attributes and improve the performance of the recommendation system, comprehensive consideration of UIG, ICG and SNEG can more intuitively show the consideration of various factors when modeling, as shown in the attached figure. Figure 2 As shown in Figure 2, the user-item global interaction category graph formed by UIG, ICG, and SNEG is not a fused graph, but is only used as an intuitive representation.

[0154] User potential representation of item category information model Include users' potential representation in the item space in building item category information models and users’ latent representation in social space for modeling item category information ; The formula for obtaining is:

[0155] ;

[0156] in, The potential representation of the project space in the project category information model is acquired through graph neural network learning. Acquiring latent representations of social space in modeling item category information through graph neural network learning.

[0157] The formula for obtaining is:

[0158] ;

[0159] In the formula, Indicates user Interactive Project Latent representation in item space;

[0160] in, The formula for obtaining is:

[0161] ;

[0162] ;

[0163] ;

[0164] In the formula, Indicates user Interactive Project The initial embedding representation of the category; Indicates user About Project The initial embedding representation of the rating is Indicates user About Project Initial representation of category preferences;

[0165] Introduce the attention mechanism in the aggregation function:

[0166] ;

[0167] in, represents the weights of different interaction items for user feature representation after normalization using the Softmax function, and:

[0168] ;

[0169] .

[0170] The formula for obtaining is:

[0171] ;

[0172] in, Indicates user The first-order neighbor nodes Item purchase preference embedding;

[0173] After introducing the attention mechanism:

[0174] ;

[0175] In the formula, represents the weights of different first-order neighbor users for user feature representation after normalization using the Softmax function, and:

[0176] ;

[0177] .

[0178] Project Potential Representation of Project Category Information Model Learning latent representation acquisition of user space via graph neural networks. The formula for obtaining is:

[0179] ;

[0180] In the formula, Display items A latent representation of the interacting user in user space;

[0181] After introducing the attention mechanism:

[0182] ;

[0183] In the formula, represents the weights of the potential representations of different users to items after normalization using the Softmax function, and:

[0184] ;

[0185] .

[0186] Step S03) and step S04) contain several initial embedding representations, which are all generated by nn.Embedding. nn.Embedding.weight will be initialized from the standard normal distribution into a matrix of size (num_embeddings, embedding_dim). nn.Embedding is updated through back propagation during training. Its parameters are a matrix, each row represents the word vector of a word, and each column represents the dimension of a word vector. When the model predicts a different result from the correct result during training, the loss function is calculated, and then the gradient is calculated through back propagation, and then the parameters in the model are updated to minimize the loss function.

[0187] S04) User-item global information fusion;

[0188] The final embedding representation of the user is obtained by fusing the social importance model and the user potential representation in the item category information through a multi-layer perceptron. ;

[0189] The final embedding representation of the project is obtained by fusing the social importance model and the potential representation of the project in the project category information through a multi-layer perceptron ;

[0190] S05) predicting the user's ratings for uninteracted items;

[0191] Splicing and , input the concatenated representation into the multi-layer perceptron to obtain the user About Project The true rating value ;

[0192] The user's rating of the non-interacted items is obtained by training the prediction model, and the objective function of the prediction model is:

[0193] ;

[0194] In the formula, Represents the training sample set the number of Indicates user About Project The predicted score value of Indicates user About Project The actual rating value.

[0195] In this embodiment, RMSprop in the literature Tijmen Tieleman and Geoffrey Hinton. 2012.Lecture 6.5-RMSProp, COURSERA: Neural networks for machine learning.University of Toronto, Technical Report (2012) is selected as the objective function optimizer. There are 5 random initial embeddings in the model training, namely user initial embedding , the project is initially embedded , rating category embedding , user social network influence weight embedding and project type embedding ,In order to avoid the overfitting problem of deep learning, a dropout strategy is adopted in the training phase, but this strategy needs to be disabled in the testing phase in order to test the entire network.

[0196] The trained model can predict the user's ratings for uninteracted items, and after sorting the items, it can determine the items recommended for the user.

[0197] The above is only an embodiment of the present application and is not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application should be included in the scope of the claims of the present application.

Claims

1. A method for optimizing the performance of a social recommendation system based on social importance and item category information, characterized in that: The method is implemented based on a social network graph, a user-project graph, and a project-category graph. The social network graph is used to describe the social relationship between users, the user-project graph is used to describe the interactive relationship between users and projects, and the project-category graph is used to describe the category to which the project belongs. The method specifically includes the following steps: S01) Establish similar user models for cold start users; The project purchase record of each user in the social network graph is obtained, and the cold start user is defined based on the number of interactions between the user and its own node and project; users with the same purchase preference as the cold start user are obtained based on the historical interaction data of the user and the project, and the users with the same purchase preference are used as similar users of the cold start user; based on the social network graph of the cold start user and the similar users of the cold start user, a social network enhanced graph is generated; in the social network enhanced graph, the first-order neighbor users of the user constitute the social space of the user, the first-order projects of the user constitute the project space of the user, and the first-order users of the project constitute the user space of the project; S02) Build a social importance model; The PageRank value of each user in the social network is P The integer part of the product of is taken as the user evaluation weight, which is used to characterize the importance of the user in the social network; P is the smallest positive integer that makes the product greater than 1; for users without social information, their user evaluation weight is 1; Latent representation of users in social importance model , including the latent representation of users in the item space in building social importance models and the latent representation of users in the social space for building social importance models ; The formula for obtaining is: ； In the formula, represents a multi-layer perceptron, Represents a vector concatenation operation; in, The latent representation of the item space in the social importance model is acquired through graph neural network learning. Acquire the latent representation of social space in social importance model through graph neural network learning; Latent representation of items in social importance model , the latent representation acquisition of user space in social importance model is established through graph neural network learning; S03) Establish project category information model; User potential representation of item category information model , including the potential representation of users in the item space in building item category information models and users’ latent representation in social space for modeling item category information ; The formula for obtaining is: ； in, The potential representation of the project space in the project category information model is acquired through graph neural network learning. Acquire the latent representation of social space in the project category information model through graph neural network learning; Project Potential Representation of Project Category Information Model Acquire the latent representation of the user space in the project category information model through graph neural network learning; S04) User-item global information fusion; The final embedding representation of the user is obtained by fusing the social importance model and the user potential representation in the item category information through a multi-layer perceptron. ; The final embedding representation of the project is obtained by fusing the social importance model and the potential representation of the project in the project category information through a multi-layer perceptron ; S05) predicting the user's ratings for uninteracted items; Splicing and , input the concatenated representation into the multi-layer perceptron to obtain the user About Project The predicted score value of ; The user's rating of the non-interacted items is obtained by training the prediction model, and the objective function of the prediction model is: ； In the formula, Represents the training sample set the number of Indicates user About Project The predicted score value of Indicates user About Project The actual rating value.

2. A social recommendation system performance optimization method based on social importance and item category information according to claim 1, characterized in that: In step S01), a cold start user is defined as a user whose number of interactions with its neighboring nodes and projects is less than K The next node, K Is a positive integer.

3. The method for optimizing the performance of a social recommendation system based on social importance and item category information according to claim 1, characterized in that: In step S02), The acquisition method is: Learning latent representations of items in item space via graph neural networks , aggregate using aggregate functions , and introduce an attention mechanism in the aggregation function to obtain the weights of different items for user feature representation, where the user feature is a multi-dimensional representation of the user embedding; The specific formula is: ； In the formula, Represents the Relu activation function; Represent the weights and biases of the neural network respectively; Indicates user Interactive Project Latent representation in item space, The dimension is d A real vector of ; Represents an aggregate function, which is an addition function; Indicates user Interactive Project Collection the number of in, ; ； In the formula, Indicates user Interactive Project The initial embedding representation of Indicates user The initial embedding representation of user rating weights, Indicates user Interactive Project The initial embedding representation of the evaluation; Indicates user The initial embedding representation obtained based on the status and ratings in the social network, Represents the Tensor product symbol of two vectors; After introducing the attention mechanism: ； In the formula, represents the weights of different interaction items for user feature representation after normalization using the Softmax function; and: ； ； in, Indicates user The initial embedding representation of represents the weight of the neural network, Represents the bias of the neural network.

4. The method for optimizing the performance of a social recommendation system based on social importance and item category information according to claim 3, characterized in that: In step S02), The method to obtain is: Learning latent representations of first-order neighbor users in item space via graph neural networks ,by As the first-order neighbor user-item purchase preference embedding representation, use the aggregation function to aggregate , and introduces an attention mechanism in the aggregation function to obtain the weights of different first-order neighbor users for user feature representation; The specific formula is: ； In the formula, Indicates user The first-order neighbor nodes Item purchase preference embedding representation; After introducing the attention mechanism: ； In the formula, represents the weights of different first-order neighbor users for user feature representation after normalization using the Softmax function, and: ； 。 5. A method for optimizing the performance of a social recommendation system based on social importance and item category information according to claim 4, characterized in that: In step S02), The method to obtain is: Learning latent representations of users interacting with items in user space via graph neural networks , aggregate using aggregate functions , and introduce the attention mechanism in the aggregation function to obtain the weights of different users' potential representations of items; the specific formula is: ； In the formula, Display items A latent representation of the interacting user in user space; Display items Interactive user collection the number of Represents an aggregate function, which is an addition function; in, The method to obtain is: ； ； In the formula, Indicates user The initial embedding representation of Indicates user In the initial embedding representation of user rating weights, Indicates user About Project The initial embedding representation of the evaluation; After introducing the attention mechanism: ； In the formula, represents the weights of different users' potential representations of items after normalization using the Softmax function, and: ； 。 6. A method for optimizing the performance of a social recommendation system based on social importance and item category information according to claim 5, characterized in that: In step S03), The formula for obtaining is: ； In the formula, Indicates user Interactive Project Latent representation in item space; in, The formula for obtaining is: ； ； ； In the formula, Indicates user Interactive Project The initial embedding representation of the category; Indicates user About Project The initial embedding representation of the rating is Indicates user About Project Initial embedding representation of category preferences; Introduce the attention mechanism in the aggregation function: ； in, represents the weights of different interaction items for user feature representation after normalization using the Softmax function, and: ； 。 7. A method for optimizing the performance of a social recommendation system based on social importance and item category information according to claim 6, characterized in that: In step S03), The formula for obtaining is: ； in, Indicates user The first-order neighbor nodes Item purchase preference embedding; After introducing the attention mechanism: ； In the formula, represents the weights of different first-order neighbor users for user feature representation after normalization using the Softmax function, and: ； 。 8. The method for optimizing the performance of a social recommendation system based on social importance and item category information according to claim 7, characterized in that: In step S03), The formula for obtaining is: ； In the formula, Display items A latent representation of the interacting user in user space; After introducing the attention mechanism: ； In the formula, represents the weights of the potential representations of different users to items after normalization using the Softmax function, and: ； 。

Images