A graph learning-based attribute network social recommendation system and method

By combining graph learning and social aggregation neural network models with the topological structure and attribute information of users and items, high-order neighbor information is captured, which solves the problem of insufficient accuracy and stability of existing social recommendation models on large-scale networks and achieves high-quality recommendation results.

CN118227895BActive Publication Date: 2024-11-08CHANGSHU INSTITUTE OF TECHNOLOGY
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202410556814.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2024-05-07
Publication Date
2024-11-08
Estimated Expiration
2044-05-07

AI Technical Summary

Technical Problem

Existing social recommendation models fail to effectively utilize graph learning techniques, resulting in insufficient accuracy and stability of recommendations on large-scale networks, and they also ignore high-order neighbor information between users and items.

Method used

We employ a graph learning-based attribute network social recommendation method. By combining graph representation learning and social aggregation neural network models with the topological structure and attribute information of users and items, we use multilayer perceptrons and attention mechanisms to capture the latent embeddings of users and items, and perform high-order neighbor information fusion and diffusion to predict user-item preference scores.

Benefits of technology

It improves the accuracy and stability of social recommendations, enables high-quality recommendations on large-scale networks, preserves data integrity, solves the data sparsity problem, and enhances the accuracy of recommendations.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118227895B_ABST
    Figure CN118227895B_ABST
Patent Text Reader

Abstract

The present invention discloses a social recommendation method for attribute network based on graph learning, which comprises the following steps: embedding free user vectors and free item vectors to obtain an initial embedding layer, using the initial embedding layer to convert the one-time representation of users and items into user-free embedding vectors and item-free embedding vectors, encoding the attribute information of users and items in the graph, and obtaining user- and item-related feature vectors; then fusing the embedding vectors and feature vectors to obtain the initial embedding of users and items; then capturing the potential embedding of users for input data, applying node-level weight information to the graph-level attention mechanism, and performing an update operation on users; when the update process is stable, the output layer predicts the preference score of each unobserved user-item pair, and selects the user-item with the highest score for social recommendation. The present invention can perform social recommendation on a large-scale network, and improves the accuracy and stability of social recommendation.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of network recommendation systems, and in particular to an attribute network social recommendation method based on graph learning. Background Art

[0002] In recent years, the explosive growth of data has caused information overload for users, hindering users from effectively searching for items that meet their functional needs. In this case, recommendation systems play an increasingly important role in helping users overcome data or service overload and automatically recommend appropriate data or services. Recommendation systems capture users' preference information and recommend items that may be of interest to them. They can provide users with effective suggestions in complex service data, allowing them to quickly find effective data.

[0003] With the rise of online social networking platforms, such as Facebook and Twitter, users have established social relationships and shared their opinions and life experiences on these platforms. Thanks to the rich information sources on social networks, social recommendation has become the focus of many researchers.

[0004] Early social recommendation methods relied on leveraging the interests of each user's first-order social neighbors for better user modeling, and failed to model the social influence diffusion process from the global social network structure. In fact, since users play a central role in both the user's social behavior and the user-item interest behavior, the key to social recommendation lies in learning user embeddings with both behaviors. Due to the effects of social homogeneity and social influence, leveraging social relationships in recommendations works well. Some models utilize first-order social neighbors for recommendations, which alleviates the data sparsity problem to a certain extent. Although the performance of these social recommendation models has improved, current social recommendation models are far from satisfactory.

[0005] Most existing social recommendation methods either focus on traditional methods such as matrix analysis or collaborative filtering, which use user-item interest behavior data to learn user and item embeddings. However, most users have limited behavior data and this method suffers from data sparsity problems; or focus on feature information such as context or community detection. These surveys largely ignore methods using modern deep learning techniques, especially graph learning. Most data in recommendation systems are essentially graph-structured, and graph learning is effective in capturing the connection relationships between nodes and representation learning of graph data. It can capture the correlation information of the graph by passing messages between the nodes of the graph. Summary of the invention

[0006] The purpose of the present invention is to provide an attribute network social recommendation method based on graph learning with simple algorithm, low resource consumption and high accuracy.

[0007] The technical solution to achieve the purpose of the present invention is: a method for attribute network social recommendation based on graph learning, which specifically includes the following steps:

[0008] Step 1: According to the topological structure information and attribute information given by the attribute network, the free user vector and the free item vector are embedded to obtain the initial embedding layer;

[0009] Step 2: Use the initial embedding layer to convert the one-shot representation of user a and item i into user-free embedding vectors p a and the item-free embedding vector q i ;

[0010] Step 3: Divide the attribute information of the user and the item in the figure into digital type and text type, and encode them in different ways to obtain the user-related feature vector x a and the item-related feature vector y i ;

[0011] Step 4: embed the userless vector p a and user-related feature vector x a Use MLP to fuse and get user fusion embedding Embed the no-item vector q i and the item-related feature vector y i Use MLP to fuse and get item fusion embedding

[0012] Step 5: Build a social aggregation neural network model As the initial embedding of the social aggregation neural network model, i.e., layer 0, the attention mechanism is used on the input data to capture the user’s potential embedding and act on the next layer;

[0013] Step 6: Apply the node-level weight information to the graph-level attention mechanism, update the user, and integrate the node information of the previous layer to achieve graph-level modeling;

[0014] Step 7: The social aggregation neural network model aggregates user embeddings and performs a full connection operation on the user vector to diffuse the user vector;

[0015] Step 8: Call the user or item embedding information of each layer to use MLP interaction modeling, predict the preference score of each unobserved user-item by training the user-item embedding, select the user-item with the highest score for social recommendation, and get the final recommendation list.

[0016] A graph learning-based attribute network social recommendation system, which is used to implement the graph learning-based attribute network social recommendation method, and the system includes an input layer, an information fusion layer, a user diffusion layer, an item diffusion layer and a rating prediction layer, wherein:

[0017] The input layer outputs a low-dimensional vector by obtaining relevant inputs;

[0018] The information fusion layer fuses attribute features and free embedding;

[0019] The user diffusion layer sets a multi-level attention structure for diffusing high-order user networks;

[0020] The item diffusion layer diffuses user and item information through a graph convolutional network;

[0021] The rating prediction layer,predicts the preference scores between each unobserved user,item after the diffusion process reaches stability.

[0022] Compared with the prior art, the present invention has the following significant advantages: (1) The present invention fully utilizes the advantages of deep learning technology and uses graph representation learning to solve the social recommendation problem. It can not only perform social recommendation on large-scale networks, but also retain data in combination with the previous information to retain data integrity to the greatest extent; (2) The present invention integrates the network topology, node feature information and its neighbor node information, so that node features can be learned and expressed with higher quality, and social recommendation can be performed more efficiently. In addition, the present invention can solve large-scale social network data and improve the accuracy and stability of social recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] Figure 1 The present invention is a flowchart of an attribute network social recommendation method based on graph learning.

[0024] Figure 2 It is a schematic diagram of the process of encoding and decoding attribute information in the present invention.

[0025] Figure 3 It is a schematic diagram of the process of fusing the embedding vector and the feature vector in the present invention.

[0026] Figure 4 It is a schematic diagram of the process of performing update operation on a user in the present invention.

[0027] Figure 5 It is a schematic diagram of the process of updating the project in the present invention.

[0028] Figure 6 It is a schematic diagram of the process of predicting user-item pairs in the present invention. DETAILED DESCRIPTION

[0029] The present invention provides an attribute network social recommendation method based on graph learning, which specifically includes the following steps:

[0030] Step 1: According to the topological structure information and attribute information given by the attribute network, the free user vector and the free item vector are embedded to obtain the initial embedding layer;

[0031] Step 2: Use the initial embedding layer to convert the one-shot representation of user a and item i into user-free embedding vectors p a and the item-free embedding vector q i ;

[0032] Step 3: Divide the attribute information of the user and the item in the figure into digital type and text type, and encode them in different ways to obtain the user-related feature vector x a and the item-related feature vector y i ;

[0033] Step 4: embed the userless vector p a and user-related feature vector x a Use MLP to fuse and get user fusion embedding Embed the no-item vector q i and the item-related feature vector y i Use MLP to fuse and get item fusion embedding

[0034] Step 5: Build a social aggregation neural network model As the initial embedding of the social aggregation neural network model, i.e., layer 0, the attention mechanism is used on the input data to capture the user’s potential embedding and act on the next layer;

[0035] Step 6: Apply the node-level weight information to the graph-level attention mechanism, update the user, and integrate the node information of the previous layer to achieve graph-level modeling;

[0036] Step 7: The social aggregation neural network model aggregates user embeddings and performs a full connection operation on the user vector to diffuse the user vector;

[0037] Step 8: Call the user or item embedding information of each layer to use MLP interaction modeling, predict the preference score of each unobserved user-item by training the user-item embedding, select the user-item with the highest score for social recommendation, and get the final recommendation list.

[0038] As a specific example, in step 2, the initial embedding layer is used to convert the one-shot representations of user a and item i into user-free embedding vectors p a and the item-free embedding vector qi , as follows:

[0039] Use the pre-trained initial embedding layer to map each user and item into a low-dimensional embedding representation and set it as the user-free latent embedding matrix and the item-free latent embedding matrix as the user and item vector lookup table. Perform the index operation by querying the vector lookup table to convert the one-time representation of user a and item i into the user-free embedding vector p a and the item-free embedding vector q i .

[0040] As a specific example, in step 3, the attribute information of the user and the attribute information of the item in the figure are divided into digital type and text type, and encoded in different ways to obtain the user-related feature vector x a and the item-related feature vector y i , as follows:

[0041] The user attribute information and project attribute information in the graph are divided into digital type and text type, and then encoded in different ways. For discrete attributes, one-hot encoding is used, and the one-hot encoding is multiplied with the randomly initialized transformation matrix to obtain the initial embedding; for continuous attributes similar to text, word vector Word2Vec is used for encoding, and the Word2Vec encoding is multiplied with the randomly initialized transformation matrix to obtain the initial embedding.

[0042] As a specific example, in step 4, the user-free embedding vector p a and user-related feature vector x a Use MLP to fuse and get user fusion embedding Embed the no-item vector q i and the item-related feature vector y i Use MLP to fuse and get item fusion embedding The details are as follows:

[0043] Embed the userless vector p a and user-related feature vector x a Use MLP for fusion, the fusion method is:

[0044]

[0045] where g u (x) is the transformation function; It is a vector concatenation operator. This fusion method spreads from the inner layer to the outer layer.

[0046] Embed the no-item vector q i and the item-related feature vector y iUse MLP for fusion, the fusion method is:

[0047]

[0048] Among them, g v (x) is the transformation matrix.

[0049] As a specific example, in step 5, an attention mechanism is used on the input data to capture the user's potential embedding, as follows:

[0050]

[0051] Where S a is the user set for user a, b is the neighbor node acting on user a, The node attention network representing user b to user a in the k+1th layer of the social network is a multi-layer perceptron MLP to learn the node weights:

[0052]

[0053] Social influence strength takes the embeddings of the two related users at the kth layer as input and sends these features to the MLP, represented as NodeAtt to learn the relationship between the features for social influence strength learning, which ultimately plays a role in balancing the weights of different adjacent nodes in the graph for each user.

[0054] As a specific example, in step 6, the node-level weight information is applied to the graph-level attention mechanism to update the user, and the node information of the previous layer is integrated to realize graph-level modeling, as follows:

[0055]

[0056] in, It is a multi-graph attention network that learns to fuse and aggregate information from different graphs, and a multi-layer perceptron MLP, denoted as MuitiGraphAtt1, to learn graph-level weights:

[0057]

[0058] Graph-level weights balance social influence and history for each user for user embedding modeling.

[0059] As a specific example, in step 7, since the project information is jointly influenced by the information between users, the social aggregation information is aggregated into the project module for joint use, and the project embedding has static characteristics. Diffusion of user information and project information will make the result more accurate and the information coverage more comprehensive, so that it can be diffused to a greater extent. We use To learn graph-level latent embeddings for the user-item diffusion process:

[0060]

[0061] Use multi-graph attention mechanism learning, also multi-layer perceptron MLP, namely MuitiGraphAtt2, to learn graph-level weights.

[0062] As a specific example, in step 8, when the update process reaches stability, the output layer predicts the preference score of each unobserved user-item pair, specifically:

[0063]

[0064] z2=σ([W2·z1])

[0065] …

[0066] z l-1 =σ([W l-1 ,·z l-2 ])

[0067]

[0068] Where W is the weight matrix, which is continuously updated during the training process, l is the number of hidden layers; σ is a nonlinear activation function, using ReLU or Tanh, is the final prediction result;

[0069] During the update process, a pairwise sorting-based loss function Adam is used for optimization. This function is widely used in implicit feedback, so that the model can better fit the data and make the prediction results more accurate:

[0070]

[0071] Where R + represents the positive sample set, that is, the set of items that the user actually interacts with; R - represents the negative sample set, that is, the set of items that have not interacted with the user through random samples, represents the probability of the user predicting item i in the positive sample set, represents the probability of item j in the negative sample set; σ() is the activation function, and the Sigmoid function or the LeakyReLU function is selected; λ is the regularization coefficient that controls the complexity of item and user embedding. are the trainable parameters used.

[0072] As a specific example, in steps 5 to 7, the model is built by using the TensorFlow framework, which uses the Adam optimizer to optimize all models, with a batch size of 512, an embedding size of {16, 32, 64}, and a learning rate search in [0.0005, 0.001, 0.005, 0.01, 0.05, 0.1]; Gaussian distribution is used to randomly initialize no user embeddings and no item embeddings as well as weight parameters, where the mean and standard deviation distributions of all models are set to 0 and 0.1.

[0073] The present invention also provides an attribute network social recommendation system based on graph learning, which is used to implement the attribute network social recommendation method based on graph learning. The system includes an input layer, an information fusion layer, a user diffusion layer, an item diffusion layer and a rating prediction layer, wherein:

[0074] The input layer outputs a low-dimensional vector by obtaining relevant inputs;

[0075] The information fusion layer fuses attribute features and free embedding;

[0076] The user diffusion layer sets a multi-level attention structure for diffusing high-order user networks;

[0077] The item diffusion layer diffuses user and item information through a graph convolutional network;

[0078] The rating prediction layer,predicts the preference scores between each unobserved user,item after the diffusion process reaches stability.

[0079] The main contributions of this invention include three aspects:

[0080] 1) A high-level data social recommendation method that integrates graph neural networks and attention mechanisms is proposed to provide a list of user recommended items by analyzing the data in the network;

[0081] 2) Apply graph neural networks in deep learning to process graph structures, capture the connection relationship between nodes and learn the representation of graph data;

[0082] 3) We focus on the influence of neighbor users and propose the influence of higher-order neighbors on users based on the first-order social neighbors, thereby integrating the information of first-order neighbors and more distant neighbors in the graph to jointly construct a social recommendation model.

[0083] The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

[0084] Example

[0085] Combination Figure 1The present invention provides an attribute network social recommendation method based on graph learning, comprising the following steps:

[0086] Step 1: According to the topological structure information and attribute information given by the attribute network, the free user vector and the free item vector are embedded to obtain the initial embedding layer;

[0087] Step 2: Use the pre-trained initial embedding layer to convert the one-shot representation of user a and item i into a user-free embedding vector p a and the item-free embedding vector q i , as follows:

[0088] Use the pre-trained initial embedding layer to map each user and item into a low-dimensional embedding representation and set it as the user-free latent embedding matrix and the item-free latent embedding matrix as the user and item vector lookup table. Perform an index operation by querying the vector lookup table to convert the one-shot representation of user a and item i into the user-free embedding vector p a and the item-free embedding vector q i .

[0089] Step 3: Divide the attribute information of the user and the item in the figure into digital type and text type, encode them in different ways, and obtain the user-related feature vector x a and the item-related feature vector y i , as follows:

[0090] Traditional graph machine learning requires the use of artificial feature engineering to achieve the conversion of graphs to D-dimensional vectors, while graph representation learning processes graphs by automatically learning features to convert each modal input into a vector. This process does not require human intervention. The graph is input into the encoder for encoding, and then put into the decoder to obtain the corresponding vector for dot product to obtain the cosine similarity, such as Figure 2 shown.

[0091] For the attribute information of users and projects, discrete information such as ID, age, price, etc. is distinguished from discrete information such as notes and comments, and they are processed separately to obtain encoding; for discrete attribute information such as ID, age and price, one-hot encoding is performed, and the initial embedding can be obtained by multiplying one-hot with a randomly initialized transformation matrix; for continuous attribute information such as notes and comments that are similar to text, word vectors such as Word2Vec are used for corresponding embedding to obtain the output information of the initial embedding layer.

[0092] Step 4: embed the userless vector p a , the non-item embedding vector q i and user-related feature vector x a , item-related feature vector y iUse MLP for fusion to get the initial embedding of users and items at layer 0 and Combination Figure 3 , as follows:

[0093] Embed the userless vector p a and user-related feature vector x a Use MLP for fusion, the fusion method is:

[0094]

[0095] where g u (x) is the transformation function, which is set to the unit function to generalize many classic fusion operations. It is a vector concatenation operator. This fusion method spreads from the inner layer to the outer layer.

[0096] Embed the no-item vector q i and the item-related feature vector y i Use MLP for fusion, the fusion method is:

[0097]

[0098] Similarly, g v (x) is also a transformation matrix, is the vector concatenation operator.

[0099] Step 5: Use the attention mechanism to learn the impact of global social interaction on users in social interactions, including the impact from first-order neighbors and higher-order neighbors. In order to better learn user embedding, according to the dynamic propagation and aggregation process of social information, the idea of ​​the attention mechanism is used to capture the user's potential embedding for the input data, and act on the next layer to prevent excessive data loss. Figure 4 , as follows:

[0100]

[0101] Where S a is the user set for user a, b is the neighbor node acting on user a, Represents the node attention network of user b to a at the k+1th layer in the social network, which is a multi-layer perceptron (MLP) to learn the node weights:

[0102]

[0103] Social influence strength takes the embeddings of the two related users at the kth layer as input and sends these features to the MLP, which can be represented as NodeAtt to learn the complex relationship between features for social influence strength learning (alignment and normalization are required), and ultimately plays a role in balancing the weights of different adjacent nodes in the graph for each user.

[0104] Step 6: Apply the node-level weight information to the graph-level attention mechanism, update the user, and integrate the node information of the previous layer to achieve graph-level modeling, as follows:

[0105]

[0106] in, It is a multi-graph attention network that learns to fuse and aggregate information from different graphs. They are also multi-layer perceptrons (MLP) represented here as MuitiGraphAtt to learn graph-level weights:

[0107]

[0108] Graph-level weights balance social influence and history for each user for user embedding modeling.

[0109] Furthermore, since project information is jointly influenced by information between users, we aggregate social aggregate information into the project module for joint use. We think this is very reasonable. Project embedding has static characteristics. Diffusion of user information and project information will make the results more accurate and the information coverage more comprehensive, so that it can be diffused to a greater extent. We use To learn the graph-level latent embedding of the user-item diffusion process, such as Figure 4 , Figure 5 As shown:

[0110]

[0111] Here, the multi-graph attention mechanism is used to learn the graph-level weights using a multi-layer perceptron MLP, namely MuitiGraphAtt.

[0112] Step 7: When the update process reaches stability, the output layer predicts the preference score of each unobserved user-item pair, selects the user-item with the highest score for social recommendation, and obtains the final recommendation list. Figure 6 , as follows:

[0113] When the updating process reaches stability, the output layer predicts the preference score for each unobserved user-item pair as follows:

[0114]

[0115] z2=σ([W2·z1]),

[0116] …

[0117] z l-1 =σ([W l-1 ,·z l-2 ]),

[0118]

[0119] Where W is the weight matrix, which can be continuously updated during the training process, l is the number of hidden layers, and σ is the nonlinear activation function, which can use ReLU or Tanh.

[0120] Furthermore, the update process requires the optimization of a pairwise ranking-based loss function, which is widely used in implicit feedback:

[0121]

[0122] Where R + represents the positive sample set (the set of items that the user actually interacts with), R - represents the negative sample set (the set of items that have not interacted with the user through random samples), represents the probability of the user predicting item i in the positive sample set, represents the probability of item j in the negative sample set, σ() is the activation function, which can use Sigmoid function, Leaky ReLU function, etc., λ is the regularization coefficient that controls the complexity of item and user embedding, are the trainable parameters used.

[0123] Furthermore, we implement our proposed model by using TensorFlow framework, which uses Adam optimizer to optimize all models with a batch size of 512. The embedding size is {16, 32, 64}, and the learning rate is searched in [0.0005, 0.001, 0.005, 0.01, 0.05, 0.1]. We randomly initialize the no-user embedding and no-item embedding as well as the weight parameters using Gaussian distribution, where the mean and standard deviation of all models are set to 0 and 0.1.

[0124] In summary, the attribute network social recommendation method based on graph learning of the present invention builds user and project models based on the data of the attribute network, aggregates the information of neighbor nodes from the first order and higher order, and combines the aggregated neighbor information with the self information and iteratively updates to obtain the latest representation of the node. Finally, according to the updated node embedding of each layer, the predicted value of the unclassified node is obtained through model training and prediction, and the case with the highest probability is selected for social recommendation to obtain the final recommendation list.

[0125] The present invention makes full use of the advantages of deep learning technology and uses graph representation learning to solve the social recommendation problem. It can not only perform social recommendations on large-scale networks, but also retain data in combination with the previous information to retain data integrity to the greatest extent. The present invention integrates the network's topological structure, node feature information, and its neighbor node information, so that node features can be learned and expressed with higher quality, and social recommendations can be performed more efficiently. In addition, the present invention can solve large-scale social network data and improve the accuracy and stability of social recommendations.

[0126] The above are only preferred embodiments of the present invention. It should be pointed out that, for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A graph learning-based attribute network social recommendation method, characterized in that: The specific steps include: Step 1: According to the topological structure information and attribute information given by the attribute network, the free user vector and the free item vector are embedded to obtain the initial embedding layer; Step 2: Use the initial embedding layer to convert the one-shot representation of user a and item i into user-free embedding vectors p a and the item-free embedding vector q i ; Step 3: Divide the attribute information of the user and the item in the figure into digital type and text type, and encode them in different ways to obtain the user-related feature vector x a and the item-related feature vector y i ; Step 4: embed the userless vector p a and user-related feature vector x a Use MLP to fuse and get user fusion embedding Embed the no items into vector q i and the item-related feature vector y i Use MLP to fuse and get item fusion embedding Step 5: Build a social aggregation neural network model As the initial embedding of the social aggregation neural network model, i.e., layer 0, the attention mechanism is used on the input data to capture the user’s potential embedding and act on the next layer; Step 6: Apply the node-level weight information to the graph-level attention mechanism, update the user, and integrate the node information of the previous layer to achieve graph-level modeling; Step 7: The social aggregation neural network model aggregates user embeddings and performs a full connection operation on the user vector to diffuse the user vector; Step 8: Call the user or item embedding information of each layer to use MLP interaction modeling, predict the preference score of each unobserved user-item by training the user-item embedding, select the user-item with the highest score for social recommendation, and obtain the final recommendation list; In step 5, an attention mechanism is used on the input data to capture the user’s potential embedding, as follows: Where S a is the user set for user a, b is the neighbor node acting on user a, The node attention network representing user b to user a in the k+1th layer of the social network is a multi-layer perceptron MLP to learn the node weights: Social influence strength takes the embeddings of the two related users at the kth layer as input and sends these features to the MLP, represented as NodeAtt to learn the relationship between the features for social influence strength learning, which ultimately plays a role in balancing the weights of different adjacent nodes in the graph for each user; In step 6, the node-level weight information is applied to the graph-level attention mechanism to update the user and integrate the node information of the previous layer to achieve graph-level modeling, as follows: in, It is a multi-graph attention network that learns to fuse and aggregate information from different graphs, and a multi-layer perceptron MLP, denoted as MuitiGraphAtt1, to learn graph-level weights: Graph-level weights balance social influence and history for each user for user embedding modeling; In step 7, since the project information is jointly influenced by the information between users, the social aggregation information is aggregated into the project module for joint use, and the project embedding has static characteristics, so the user information and project information are diffused together, using To learn graph-level latent embeddings for the user-item diffusion process: Use multi-graph attention mechanism learning, also multi-layer perceptron MLP, namely MuitiGraphAtt2, to learn graph-level weights.

2. The attribute network social recommendation method based on graph learning according to claim 1, characterized in that: In step 2, the initial embedding layer is used to convert the one-shot representations of user a and item i into user-free embedding vectors p respectively. a and the item-free embedding vector q i , as follows: Use the pre-trained initial embedding layer to map each user and item into a low-dimensional embedding representation and set it as the user-free latent embedding matrix and the item-free latent embedding matrix as the user and item vector lookup table. Perform the index operation by querying the vector lookup table to convert the one-time representation of user a and item i into the user-free embedding vector p a and the item-free embedding vector q i .

3. The attribute network social recommendation method based on graph learning according to claim 1, characterized in that: In step 3, the attribute information of the user and the attribute information of the item in the figure are divided into digital type and text type, and encoded in different ways to obtain the user-related feature vector x a and the item-related feature vector y i , as follows: The user attribute information and project attribute information in the graph are divided into digital type and text type, and then encoded in different ways. For discrete attributes, one-hot encoding is used, and the one-hot encoding is multiplied with the randomly initialized transformation matrix to obtain the initial embedding; for continuous attributes similar to text, word vector Word2Vec is used for encoding, and the Word2Vec encoding is multiplied with the randomly initialized transformation matrix to obtain the initial embedding.

4. The attribute network social recommendation method based on graph learning according to claim 1, characterized in that: In step 4, the user-free embedding vector p a and user-related feature vector x a Use MLP to fuse and get user fusion embedding Embed the no items into vector q i and the item-related feature vector y i Use MLP to fuse and get item fusion embedding The details are as follows: Embed the userless vector p a and user-related feature vector x a Use MLP for fusion, the fusion method is: where g u (x) is the transformation function; It is a vector concatenation operator. This fusion method spreads from the inner layer to the outer layer. Embed the no items into vector q i and the item-related feature vector y i Use MLP for fusion, the fusion method is: Among them, g v (x) is the transformation matrix.

5. The attribute network social recommendation method based on graph learning according to claim 1, characterized in that: In step 8, when the update process reaches stability, the output layer predicts the preference score of each unobserved user-item pair, specifically: z2=σ([W2·z1]) … With l-1 =σ([W l-1 ,·With l-2 ]) Where W l is the weight matrix, which is continuously updated during the training process, l is the number of hidden layers; σ is the nonlinear activation function, using ReLU or Tanh, is the final prediction result; During the update process, a pairwise sorting-based loss function Adam is used for optimization. This function is widely used in implicit feedback, so that the model can better fit the data and make the prediction results more accurate: Where R + represents the positive sample set, that is, the set of items that the user actually interacts with; R - represents the negative sample set, that is, the set of items that have not interacted with the user through random samples, represents the probability of the user predicting item i in the positive sample set, represents the probability of item j in the negative sample set; σ() is the activation function, and the Sigmoid function or the Leaky ReLU function is selected; λ is the regularization coefficient that controls the complexity of item and user embedding. are the trainable parameters used.

6. The attribute network social recommendation method based on graph learning according to claim 1, characterized in that: In steps 5 to 7, the model is built using the TensorFlow framework, which uses the Adam optimizer to optimize all models, with a batch size of 512, an embedding size of {16, 32, 64}, and a learning rate search in [0.0005, 0.001, 0.005, 0.01, 0.05, 0.1]. Gaussian distribution is used to randomly initialize the no-user embedding and no-item embedding as well as the weight parameters, where the mean and standard deviation distributions of all models are set to 0 and 0.

1.

7. An attribute network social recommendation system based on graph learning, characterized in that: The system is used to implement the attribute network social recommendation method based on graph learning according to any one of claims 1 to 6, and the system includes an input layer, an information fusion layer, a user diffusion layer, an item diffusion layer and a rating prediction layer, wherein: The input layer outputs a low-dimensional vector by obtaining relevant inputs; The information fusion layer fuses attribute features and free embedding; The user diffusion layer sets a multi-level attention structure for diffusing high-order user networks; The item diffusion layer diffuses user and item information through a graph convolutional network; The rating prediction layer,predicts the preference scores between each unobserved user,item after the diffusion process reaches stability.