A recommendation method based on graph twin network

Abstract

A recommendation method based on a graph twin network is applied to the field of personalized recommendation. Existing methods (1) lack knowledge scalability, for example, it is difficult to effectively integrate user social relationship information; (2) under the multi-layer feature information propagation paradigm, the learned features will have over-smoothing problems. Therefore, the present invention proposes a recommendation method based on a graph twin network, which models the user relationship graph and the item relationship graph through the interactive information of users and items, and mines the user relationship information and the item relationship information in the form of two isomorphic directed graphs through the graph convolution layer designed by the present invention. Finally, the user features and item features of the two channels are aggregated through the graph interaction layer to fully extract user preference information and item attribute information. The present invention effectively maintains the U‑I feature characteristics, significantly improves the accuracy of personalized recommendations, has good model scalability, and has broad application prospects.

Description

Technical Field

[0001] The present invention is applied to the field of recommendation systems based on UI relationships, and specifically involves data mining and deep learning technologies such as graph convolutional neural networks, attention mechanisms, user preference information and item attribute information feature extraction, and UI interaction information modeling. Background Art

[0002] Personalized recommendation is a comprehensive analysis task, which is widely used in social networks, music radio, e-commerce, personalized advertising, movie and video websites, and has attracted much attention. In recent years, deep learning has achieved great success in many research fields such as computer vision and natural language processing, which has aroused great interest. Deep learning is widely used in recommendation systems. Its effectiveness has been proven in the fields of information retrieval and recommendation system research, and it is the future development direction of recommendation systems. However, traditional deep learning methods have shortcomings in complexity and scalability. Model training often takes a long time. How to balance model scalability and complexity is still a big problem. At the same time, traditional deep learning is often used as a feature extraction tool in the recommendation field. It is difficult to achieve end-to-end recommendation implementation of recommendation systems by relying solely on traditional deep learning models.

[0003] The graph neural network technology was first proposed by Joan Bruna in 2014. Studies have found that graph neural networks can accurately model data in non-Euclidean space and effectively capture the internal dependencies of data, opening up new horizons for the analysis and processing of non-Euclidean data. In recent years, the research on graph neural networks has made great progress. Thomas Kpif proposed the concept of graph convolutional neural network in 2017, which provides a new idea for the processing of graph structured data and applies convolutional neural networks in deep learning to graph structured data. At the same time, Hamilton proposed an inductive learning method suitable for large-scale networks, which can quickly generate node features for new nodes without additional training, greatly alleviating the scalability problem of graph convolutional neural networks. In the research on recommendation systems, Rianne van den Berg proposed the GC-MC graph neural network recommendation framework in 2017. In 2018, Pinterest and Stanford jointly proposed PinSage, a recommendation system model based on graph convolutional neural networks. For the first time, graph convolutional neural networks were successfully applied to the industry's recommendation system, providing valuable reference and experience for the application of graph convolutional neural networks. In 2019, Wang Xiang, He Xiangnan and others proposed the NGCF recommendation method, which uses graph convolutional networks to effectively mine UI interaction information. However, most of the above methods model user item information into a single graph, which makes it difficult to effectively control the complexity of the model and the model training time. Moreover, the recommendation method based on a single graph lacks knowledge scalability and is difficult to aggregate other information sources such as multi-platform UI relationships and user social relationships.

[0004] Attention mechanism is a hot topic that is widely discussed at present. In recent years, it has been widely used in various fields of deep learning. Whether it is machine vision, natural language processing and other different types of tasks, it is easy to encounter the presence of attention models. The attention mechanism has outstanding performance in application scenarios such as modeling position weights, visual weights, and attribute weights. In the field of recommendation systems, the attention mechanism is an indispensable component for aggregating user preference information and item attribute information. In particular, the low-complexity combination of multi-layer neural networks and softmax can effectively extract the information weights of users and items while ensuring complexity. Summary of the invention

[0005] In order to realize the personalized recommendation system for users, a personalized recommendation scheme based on graph twin network is proposed. Figure 1As shown. This method uses the user relationship graph and the item relationship graph as input data. Each module fully mines the user relationship and item relationship information, effectively models the UI interaction relationship, and finally outputs the target user's predicted score for the recommended item. Specifically, this method first performs structured modeling of the user relationship graph and the item relationship graph based on the UI interaction data. Then, two layers of graph convolutional network layers designed by the present invention perform graph convolution analysis on the user relationship graph and the item relationship graph respectively, learn the node feature expression, and then effectively mine the user relationship information and the item relationship information. Then, using the graph attention mechanism, a layer of graph interaction network layer is used to model the interaction characteristics of the target user and the recommended item. The distribution law of the interaction feature vector is learned through the DNN network, and the high-order nonlinear relationship between UIs is modeled to obtain the target user's predicted score for the recommended item. Finally, the personalized recommendation task for users is realized through the top-N recommendation mechanism. The overall block diagram of this method is shown in Figure 1 shown.

[0006] The invention contents of each main module of this method are as follows:

[0007] 1. Mining user relationship information and item relationship information

[0008] The first module is the user relationship information and item relationship information mining module. The effectiveness of UI relationship information extraction directly determines the accuracy of personalized recommendation of the recommendation system. In order to fully mine user relationship and item relationship information, the present invention designs a dual-channel graph twin network to learn UI relationship features, in which each channel models and analyzes user relationship features and item relationship features through two layers of graph convolution layers. The dual-channel processing method greatly reduces the local complexity of the model while effectively improving the scalability of the model. For example, an additional edge can be added between two nodes in the user relationship graph to model the social relationship information between users.

[0009] First, based on the second-order relationship of the UI interaction network as the intermediate path, the UI interaction network is modeled as a user relationship graph and an item relationship graph. Specifically, in the UI interaction network, using items as the intermediate path, the number of second-order paths between two users is used as the edge information between the two user nodes in the user relationship graph, and then the user relationship graph is established; using users as the intermediate path, the number of second-order paths between two items is used as the edge information between the two item nodes in the item relationship graph, and then the item relationship graph is established. After the graph structured modeling is completed, the user and item relationship feature expression is modeled through two layers of graph convolutional network layers. Among them, the processing method of the two layers of graph convolutional layers is the same, and the processing method is as follows:

[0010] For any node in the current layer of the relationship graph, the graph convolution layer obtains the node feature expression of this layer by aggregating the neighbor node features of the previous layer of the node and the node feature learning. First, the self-convolution processing of all nodes in the current layer of the relationship graph is completed through the degree matrix with self-loops in the relationship graph. The advantage of this self-convolution processing method is that it can adapt to the number of node edges, thereby effectively realizing the control of the degree of node self-convolution. After the self-convolution processing is completed, the weight information of all edges of the target node is softmax normalized, and the first-order neighbor node features corresponding to the node are weighted and summed by the normalized weight information. Then, the features obtained by the self-convolution processing and the features obtained by the aggregation of the first-order neighbor relationships are added together in each dimension to preliminarily complete the node feature update of this layer. Finally, the updated node relationship features are passed through a dimensionally adjustable parameter matrix to complete the graph convolution processing of this layer. The advantage of fusing a trainable and dimensionally adjustable parameter matrix is ​​that it can effectively improve the accuracy of node feature extraction, and the low-dimensional mapping of node features can be achieved by adjusting the second dimension of the parameter matrix, which ensures the accuracy of the model while effectively reducing the training time of the model. The specific processing flow of the graph convolution layer is as follows Figure 2 shown.

[0011] 2. UI interaction feature extraction module

[0012] The second module is the UI interaction feature extraction module. The function of this module is to integrate the attention mechanism after the user and item relationship feature modeling is completed, and to model the interaction features between the target user and the recommended item through the graph interaction network layer. Among them, the attention mechanism of the present invention adopts a two-layer fully connected layer network. The distribution information of the interaction feature contains the user's preference information for the target item. The distribution regularity of the interaction feature is an important basis for realizing personalized recommendations for users.

[0013] First, for the target user node in the user relationship graph and the to-be-recommended item node in the item relationship graph, the user node features are concatenated with the to-be-recommended item node features and the features of its first-order neighbor nodes. The concatenated features are input into the self-attention network, and the output values ​​are softmax processed to obtain the weight coefficients of the item node and its first-order neighbor node features. The interaction features of the to-be-recommended item are obtained by weighted summing up the relationship features of the item node and its first-order neighbor nodes according to the obtained weights. Similarly, the item node features are concatenated with the target user node features and the features of its first-order neighbor nodes. The concatenated features are input into the self-attention network, and the output values ​​are softmax processed to obtain the weight coefficients of the user node and its first-order neighbor node features. The interaction features of the target user node and its first-order neighbor node are weighted summing up the relationship features of the target user node and its first-order neighbor nodes according to the obtained weights. Finally, the interaction features of the to-be-recommended item part and the interaction features of the target user part are concatenated to obtain the final interaction feature expression. The specific process of this module is as follows: Figure 3 shown.

[0014] 3. Interactive Inference Module

[0015] The third module is the interaction inference module. The function of this module is to learn the distribution of the interaction features to obtain the user's predicted score for the target item after the UI interaction feature extraction module extracts the interaction features between the target user and the item to be recommended. The interaction inference module is an indispensable step in building recommendations. There are many methods for designing the interaction inference module, such as the traditional feature inner product and logistic regression algorithms. However, these traditional algorithms cannot understand the feature information of users and items well. Therefore, the present invention adopts the classic DNN algorithm to learn the feature information of users and items to obtain more accurate inference results.

[0016] After obtaining the interactive feature expression of the target user and the recommended item, the interactive vector is directly input into the DNN network to obtain the initial prediction of the model. Then, the module prediction value is normalized by the sigmoid function, and the target user's predicted score for the recommended item is modeled as a preference probability expression. The module process is as follows: Figure 4 express.

[0017] 4. Top-N Recommendation Module

[0018] The last module of the present invention is the top-N recommendation module. The top-N recommendation mechanism is also the most commonly used mechanism for building a recommendation system. After obtaining the target user's rating prediction for all items in the recommendation list, all items are sorted in descending order according to the ratings, and the top N items are recommended to the user to achieve personalized recommendations for the user. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 The overall framework of the recommendation method based on graph twin network;

[0020] Figure 2 It is the framework of the user relationship information and item relationship information mining module;

[0021] Figure 3 Module framework for UI interaction feature extraction

[0022] Figure 4 Module framework for interactive speculation DETAILED DESCRIPTION

[0023] The present invention proposes a personalized recommendation method based on a graph twin network. The specific implementation steps of the invention are as follows:

[0024] Step 1: Select a public recommendation data set, number all users and items, randomly select 90% of all items that each user has interacted with as training sets, and the remaining 10% of items as test sets. Each training set and test set consists of three parts: user, item, and label. The label of the item that has interacted with the user is 1, otherwise the label is 0. The UI interaction graph is structured through all the data with the label 1 in the training set, and a connection relationship is established for users and items with interactive behaviors. Finally, according to the connection relationship of the nodes in the UI interaction graph, the user relationship graph is established by using the item as the intermediate path and the number of second-order paths between two users as the edge information between the two user nodes in the user relationship graph; the item relationship graph is established by using the user as the intermediate path and the number of second-order paths between two items as the edge information between the two item nodes in the item relationship graph.

[0025] Step 2: After completing the graph structured modeling of user relationships and item relationships, randomly initialize the node features of the two relationship graphs. Learn the node features of the two relationship graphs through the designed two-layer graph convolutional network layer, and then accurately model the user relationship information and item relationship information. The node features of any layer are obtained by aggregating the node features in the previous graph convolutional network layer and its first-order neighbor node features. The graph convolution processing method for each channel is as follows:

[0026] First, the self-convolution processing of all nodes in the relationship graph is completed through the degree matrix with self-loops in the relationship graph. Specifically, each dimension of the features of all nodes in this layer is divided by the degree of the node with self-loops. Among them, the degree matrix with self-loops is a diagonal matrix, and the diagonal elements of each dimension correspond to the degree of the node. Its mathematical expression is:

[0027]

[0028]

[0029] For example, the corresponding matrix of three nodes is expressed as follows:

[0030]

[0031] Wherein, matrix A is an adjacency matrix (square matrix), the i-th row represents the connection relationship between the i-th node of the relationship graph and other nodes, i∈{0,1,…,N}, N represents the number of nodes in the relationship graph. The j in formula (1) represents the number of columns of the adjacency matrix A, j∈{0,1,…,N}. N is the identity matrix of dimension N. The matrix represents the degree matrix with self-loops. It is a diagonal matrix. The value of the diagonal element in the i-th row represents the degree of self-loops in the i-th node of the relationship graph. After the self-convolution processing is completed, for any node N in the relationship graph i , perform softmax normalization on the weight information of the edge between the nodes, aggregate the first-order neighbor node features of the node through the normalized weight information, and then add the features after self-convolution processing and the features obtained by first-order neighbor relationship aggregation in each dimension to preliminarily complete the node feature update of this layer. The same processing is performed on all other nodes in this layer of the relationship graph. Among them, the mathematical expression of the above aggregated node information is:

[0032]

[0033] in, Represents the relationship graph K-th layer graph convolutional network layer node N i The feature expression is obtained by aggregating the feature updates of the previous layer of nodes. Represents the K-1th graph convolutional network layer node N i characteristic expression. represents the degree matrix with self-loops The value of the i-th diagonal element, that is, node N i The degree of self-loop. N(i) represents the node N i The first-order neighbor node set of node a(N a ) is any node of the set N(i). Represents the K-1th graph convolutional layer node N a The characteristic expression of Represents the node N obtained after softmax normalization a For node N i The feature update weights are The mathematical expression of is:

[0034]

[0035] Among them, L ia Represents the relationship graph node N i and its first-order neighbor node N a The value of the corresponding edge. b ) belongs to node N i Any node in the first-order neighbor node set of N(i) represents the target node N i The first-order neighbor node set of . Finally, all the updated node features are processed through a dimensionally adjustable parameter matrix to complete the graph convolution processing of this layer. Its mathematical expression is as follows:

[0036]

[0037] in, Represents the relationship graph K-th layer graph convolutional network layer node N i The feature expression is obtained by aggregating the feature updates of the previous layer nodes. σ represents a nonlinear activation function. The present invention uses the Relu activation function. W (K) Represents the parameter matrix of the Kth graph convolutional network layer. After a dimensionally adjustable parameter matrix W (K) After that, node N is completed i Convolution processing in the Kth graph convolution layer. All other nodes in this layer of the relationship graph pass through the parameter matrix W of this layer. (K) , and then complete the processing of the graph convolution layer of all nodes in this layer.

[0038] At the same time, the graph convolution layer operations of all nodes in the relationship graph can be directly completed through matrix operations. The mathematical expression of the matrix form of the graph convolution layer is as follows:

[0039]

[0040] A Lij =L ij (7)

[0041]

[0042]

[0043] Among them, function f represents the processing function of the graph convolution layer, which takes the node feature matrix H of the K-1th layer as (K-1) , the adjacency matrix A of the relationship graph, the relationship matrix A of the edge of the relationship graph L As input, the K-th layer node feature matrix H (K) is the output. Among them, the node feature matrix H (K) and H (K-1) Each row of represents the relationship feature expression of the corresponding node, A L The value of the element at any position in the matrix represents the value of the relevant edge of the relationship graph. is the edge relationship matrix A L The partial exponential treatment of A L The non-zero element values ​​of * It is a function that exponentially maps the non-zero element values ​​of the input matrix. is a matrix The degree matrix of W is a diagonal matrix. The j in formula (9) represents the number of columns of the adjacency matrix A. (K) is the parameter matrix of the Kth graph convolutional layer, W (K)The two dimensions of the parameter matrix are consistent with the node feature dimensions. σ represents the Relu nonlinear activation function.

[0044] After two layers of graph convolution were performed on the user relationship graph and the item relationship graph respectively, the user relationship information and item relationship information were fully modeled.

[0045] For example, the correlation matrix corresponding to three nodes is expressed as:

[0046]

[0047] Step 3: After the user relationship features and item relationship features are fully modeled, for the paired target user u i and items to be recommended v j , through a graph interaction layer, the relevant node relationship features are fused to obtain the corresponding interaction feature expression. First, the target user relationship features are spliced The concatenated features are input into the attention network, and the network output is processed by softmax to obtain the weight coefficients of the relationship features between the items to be recommended and their first-order neighbor nodes. The weighted sum of the corresponding item relationship features is used to obtain the recommended item v j The mathematical expression is:

[0048]

[0049]

[0050]

[0051] Among them, W1 and W2 represent the parameter matrix of the two-layer attention network, b1 and b2 represent the deviation coefficients of the two-layer attention network, and σ represents the nonlinear activation function, which uses the Relu activation function. It represents the target user u in the user relationship graph i relationship characteristics. Represents item v in the item relationship graph t The relationship characteristics of node t(v t ) belongs to the recommended items v j Any item node in its first-order neighbor item set. N(j) represents the item v to be recommended j The set of first-order neighbors of . Indicates splicing processing. is the output value of the attention network, and the input of the network is the target user relationship feature and itemsv t Relationship characteristics Then, all the output values ​​of the attention network are normalized by softmax to obtain the item node v t The corresponding weight coefficient μ it Formula (11) represents the softmax processing, and the denominator is the node v of the item to be recommended. j The sum of the attention network output values ​​corresponding to all its first-order neighbor item nodes, with the numerator being the item node v t The output value of the attention network. The denominator x represents the recommended item v j Any item node in its first-order neighbor item set. and The calculation method is the same as that of the item node x(v x ) and target user u i Relationship characteristics After concatenation, it is input into the attention network to obtain the output value of the attention network. Finally, according to the obtained item node v t The weight μ it , weighted sum of all relevant item nodes v t Relationship characteristics Get the recommended item v j Partial interaction feature z j .

[0052] Similarly, the relationship characteristics of the recommended items The concatenated features are input into the attention network, and the network output is processed by softmax to obtain the weight coefficient of the target user and its first-order neighbor node relationship features. The target user u is obtained by weighted summing the corresponding user relationship features. i The mathematical expression is:

[0053]

[0054]

[0055]

[0056] Among them, W1 and W2 represent the parameter matrix of the two-layer attention network, b1 and b2 represent the deviation coefficients of the two-layer attention network, and σ represents the nonlinear activation function, which uses the Relu activation function. Represents the item relationship graph to be recommended item node v j relationship characteristics. Represents user node u in the user relationship graph kThe relationship characteristics of node k(u k ) belongs to the target user u i Any user node in its first-order neighbor user set. N(i) represents the target user u i The set of first-order neighbor nodes. Indicates splicing processing. is the output value of the attention network, and the input of the network is the relationship feature of the items to be recommended and user u k Relationship characteristics Then, all the output values ​​of the attention network are normalized by softmax to obtain the user node u k The corresponding weight coefficient β jk Formula (14) represents the softmax processing, and the denominator is the target user node u i The sum of the attention network output values ​​corresponding to all its first-order neighbor user nodes, with the numerator being the user node u k The output value of the attention network. Among them, the denominator y represents the target user u i and any user node in its first-order neighbor user set. and The calculation method is the same as that of the user node y(u y ) Relationship features with items to be recommended After concatenation, it is input into the attention network to obtain the output value of the attention network. Finally, according to the obtained user node u k The weight β jk , weighted sum of all relevant user nodes u k Relationship characteristics Get the target user u i Partial interaction feature z i .

[0057] After getting the target user u i Partial interactive feature expression z i 、Items to be recommendedv j Partial interactive feature expression z j Finally, the interaction features of the two parts are concatenated to obtain the target user u i and items to be recommended v j The mathematical expression is:

[0058]

[0059] z i is the interaction feature of the user part, z j is the interaction feature of the item to be included. Indicates splicing processing. ij For the final target user u i And its recommended items v j The interactive feature expression of .

[0060] Step 4: After obtaining the interaction features between the target user and the recommended item, the model learns the distribution law of the interaction features through the interaction inference module, and then obtains the user u i For items v j Accurate preference prediction. Interaction inference module This method uses the classic DNN network. The DNN network can effectively learn feature distribution and model the nonlinear relationship between users and items. The interaction inference module directly inputs the interaction features into the DNN network to obtain the user's predicted score for the item, and then uses sigmoid normalization to model the model's predicted score as the user u i For items v j The preference probability expression of . Its mathematical expression is:

[0061] g1=z ij (17)

[0062] g2=σ(W1·g1+b1) (18)

[0063] g3=σ(W2·g2+b2) (19)

[0064] r' ij =sigmoid(W3·g3) (20)

[0065] Among them, z ij For target user u i And its recommended items v j W1, W2, W3 represent the parameter matrix of the DNN network, b1, b2 represent the deviation coefficients in the DNN network, σ represents the nonlinear activation function, and the Relu activation function is used. g1, g2, g3 are the interactive vector expressions of each layer of the DNN network output. r' ij It is the prediction evaluation expression of the model after sigmoid normalization, where the sigmoid in formula (20) represents the normalized mapping of the output value of the last layer of the DNN network to the sigmoid function.

[0066] Step 5: For model training, this method constructs a cross entropy loss function as the optimization target of the model, and optimizes the model's trainable parameter matrix and deviation coefficient, as well as the node features in the user-item relationship graph by minimizing the loss function value. The mathematical expression of the cross entropy loss function is:

[0067]

[0068] Where |O| represents the number of all user-item node pairs in each batch when training the model, r ij Represents user u i and itemsv j The label value of is in the range of {0,1}. A label of 0 means that item v j The attribute information does not match the user u i The preference information of user u is a negative sample. The label is 1, which means user u i With items v j There are rows with interactions, which are positive samples. For each user, we randomly sample items that the user has no interaction with as the user's training negative samples. ij Represents the prediction score of the model. By minimizing r' ij and the actual label r ij The gap is used to minimize the loss function value to complete the task of optimizing the model.

[0069] Step 6: After the model training is completed, in order to verify the effectiveness of the present invention, the algorithm of the present invention is experimented on the data sets huaban and Amazon-Book. Negative samples are randomly collected from the test set of each data set at a ratio of 1:100 to participate in the prediction and evaluation of the model. At the same time, in order to ensure the effectiveness of the model evaluation, the negative samples collected in the test set are not trained in the training set. After obtaining the model's predicted evaluation of all items for each user, the items participating in the evaluation are sorted from large to small according to the model output value, and the top N items after sorting are recommended to the target user through the top-N recommendation mechanism. The effectiveness of the method is evaluated by the recall evaluation index. Tables 1 and 2 show the comparison between the algorithm of the present invention and some cutting-edge recommendation algorithms. It can be seen that this algorithm is superior to other recommendation algorithms shown.

[0070] Table 1: Performance comparison on the Amazon-book dataset

[0071]

[0072] Table 2: Performance comparison on the Huaban dataset

[0073]

Claims

1. A recommendation method based on graph twin network, characterized in that: The following modules are applied: user relationship and item relationship information mining module, UI interaction feature extraction module, interaction inference module, and Top-N recommendation module; First, user data and item data are structured into user relationship graphs and item relationship graphs according to UI interaction relationships. The user relationship information and item relationship information mining module performs graph convolution analysis on the two relationship graphs through a dual-channel two-layer graph convolution network layer, learns the node feature expressions of the two relationship graphs, and then mines user relationship and item relationship information. The UI interaction feature extraction module uses the attention mechanism to model the interaction feature expression of the target user and the recommended item through the graph interaction network layer. The interaction inference module learns the distribution law of the interaction feature vector through the classic DNN network, models the high-order nonlinear relationship between UIs, and obtains the user's predicted score for the target item. The Top-N recommendation module implements the personalized recommendation task for the target user through the top-N recommendation mechanism. The contents of each module are as follows: 1). User relationship information and item relationship information mining module The first module is the user relationship and item relationship information mining module, which mines user and item relationship information through a dual-channel graph twin network. Each channel models and analyzes user relationship features or item relationship features through two layers of strongly principled graph convolutional network layers, thereby fully mining their relationship information. First, the second-order relationship in the UI interaction network is used as the intermediate path to model the UI interaction information as a user relationship graph and an item relationship graph. Then, a two-layer graph convolutional network layer with two channels is used to accurately model the user and item relationship feature expression. The processing method of the two-channel graph convolutional network layer is as follows: For any node in the current layer of the relationship graph, the graph convolution layer obtains the node feature expression of this layer by aggregating the neighbor node features of the previous layer of the node and the node feature learning; first, the self-convolution processing of the node in the current layer of the relationship graph is completed through the degree matrix with self-loops in the relationship graph; then, the weight information of all edges of the node is softmax normalized, and the first-order neighbor node features corresponding to the node are weighted and summed through the normalized weight information; Then, the features obtained by self-convolution and the features obtained by first-order neighbor relationship aggregation are added together in each dimension to preliminarily complete the node feature update of this layer; finally, the updated node relationship features are passed through a dimensionally adjustable parameter matrix to complete the graph convolution processing of this node in this layer; the two dimensions of this parameter matrix are consistent with the feature dimensions of the relationship graph nodes; 2). UI interaction feature extraction module The second module is the UI interaction feature extraction module. After the user-item relationship feature modeling is completed, the module uses the attention mechanism to model the interaction features between the target user and the recommended item through the graph interaction network layer. The attention mechanism uses a two-layer fully connected layer network. The processing method of the graph interaction network layer is as follows: First, for the target user node in the user relationship graph and the node of the item to be recommended in the item relationship graph, the user node features are concatenated with the node features of the item to be recommended and the features of its first-order neighbor nodes; the concatenated features are input into the self-attention network, and the output value is softmax processed to obtain the weight coefficients of the features of the item node and its first-order neighbor nodes, and the interaction features of the item node and its first-order neighbor nodes are weighted and summed according to the obtained weights to obtain the interaction features of the item to be recommended; similarly, the item node features are concatenated with the target user node features and the features of its first-order neighbor nodes; the concatenated features are input into the self-attention network, and the output value is softmax processed to obtain the weight coefficients of the features of the user node and its first-order neighbor nodes; the interaction features of the target user node and its first-order neighbor nodes are weighted and summed according to the obtained weights to obtain the interaction features of the target user part; finally, the interaction features of the item to be recommended and the interaction features of the target user part are concatenated to obtain the final interaction feature expression; 3). Interactive inference module The third module is the interactive inference module, which learns the distribution of interactive features to obtain the user's inferred score for the target item. It learns the characteristic information of users and items through the classic DNN algorithm to obtain more accurate inference results. After obtaining the interactive feature expression of the target user and the recommended item, the interactive vector is directly input into the DNN network to obtain the preliminary prediction of the model; the module prediction value is normalized by the sigmoid function, and the user's predicted score for the target item is modeled as a preference probability expression; 4). Top-N recommendation module The last module is the top-N recommendation module. After obtaining the target user's rating prediction for all items in the recommendation list, all items are sorted in descending order according to their ratings, and the top N items are recommended to the user to achieve personalized recommendations for the user. The specific implementation steps are as follows: First, we conduct graph structured modeling based on UI interaction data; establish connections between users and items with interaction records; then, based on the connection relationship of UI interaction graph nodes, we use items as intermediate paths and the number of second-order paths between two users as the edge information between the two user nodes in the user relationship graph, and then establish a user relationship graph; we use users as intermediate paths and the number of second-order paths between two items as the edge information between the two item nodes in the item relationship graph, and then establish an item relationship graph; 1). User relationship information and item relationship information mining module After completing the graph structured modeling of user relationships and item relationships; randomly initialize the node features of the two relationship graphs; learn the node features of the two relationship graphs through two layers of graph convolutional network layers, and then accurately model the user relationship information and item relationship information; the node features of any layer are obtained by aggregating the node features in the previous graph convolutional network layer and the features of its first-order neighbor nodes; the processing method of the graph convolutional layer of each channel is as follows: First, the self-convolution processing of all nodes in the relationship graph is completed through the degree matrix with self-loops in the relationship graph. Specifically, each dimension of the features of all nodes in this layer is divided by the degree of the node with self-loops. Among them, the degree matrix with self-loops is a diagonal matrix, and the diagonal elements of each dimension correspond to the degree of the node. Its mathematical expression is: Wherein, matrix A is the adjacency matrix of the relationship graph, the i-th row represents the connection relationship between the i-th node of the relationship graph and other nodes, i∈{0,1,…,N}, N represents the number of nodes in the relationship graph; j in formula (1) represents the number of columns of the adjacency matrix A, j∈{0,1,…,N}; I N is the identity matrix of dimension N; The matrix represents the degree matrix with self-loops. It is a diagonal matrix. The value of the diagonal element in the i-th row represents the degree of self-loops in the i-th node of the relationship graph. After the self-convolution processing is completed, for any node N in the relationship graph i , perform softmax normalization on the weight information of the edge between the nodes, aggregate the first-order neighbor node features of the node through the normalized weight information, and then add the features after self-convolution processing and the features obtained by first-order neighbor relationship aggregation in each dimension to preliminarily complete the node feature update of this layer; all other nodes in this layer of the relationship graph are processed in the same way; where the mathematical expression of aggregated node information is: in, Represents the relationship graph K-th layer graph convolutional network layer node N i The feature expression is obtained by aggregating the feature updates of the previous layer of nodes; Represents the K-1th graph convolutional network layer node N i characteristic expression; represents the degree matrix with self-loops The value of the i-th diagonal element, that is, node N i The degree of the self-loop; N(i) represents the node N i The first-order neighbor node set of node a(N a ) is any node of the set N(i); Represents the K-1th graph convolutional layer node N a The characteristic expression of Represents the node N obtained after softmax normalization a For node N i The feature update weights are The mathematical expression of is: Among them, L ia Represents the relationship graph node N i and its first-order neighbor node N a The value of the corresponding edge of node b(N b ) belongs to node N i Any node in the first-order neighbor node set of N(i) represents the target node N i The first-order neighbor node set of ; finally, all the updated node features are processed through a dimensionally adjustable parameter matrix to complete the graph convolution processing of this layer; its mathematical expression is as follows: in, Represents the relationship graph K-th layer graph convolutional network layer node N i The feature expression is obtained by aggregating the feature updates of the previous layer nodes; σ represents the nonlinear activation function, and the Relu activation function is used, W (K) represents the parameter matrix of the Kth graph convolutional network layer; After a dimensionally adjustable parameter matrix W (K) After that, node N is completed i Convolution processing in the Kth graph convolution layer; all other nodes in this layer of the relationship graph pass through the parameter matrix W of this layer (K) , and then complete the processing of the graph convolution layer of all nodes in this layer; At the same time, the graph convolution layer operations of all nodes in the relationship graph are completed directly through matrix operations; the mathematical expression of the matrix form processing of the graph convolution layer is as follows: A Lij =L ij (7) Among them, function f represents the processing function of the graph convolution layer, which takes the node feature matrix H of the K-1th layer as (K-1) , the adjacency matrix A of the relationship graph, the relationship matrix A of the edge of the relationship graph L As input, the K-th layer node feature matrix H (K) is the output; where the node feature matrix H (K) and H (K-1) Each row of represents the relationship feature expression of the corresponding node, A L The value of the element at any position in the matrix represents the value of the relevant edge of the relationship graph; is the edge relationship matrix A L The partial exponential treatment of A L The non-zero element values ​​of * It is a function that performs exponential mapping on the non-zero element values ​​of the input matrix; is a matrix The degree matrix of is a diagonal matrix. The j in formula (9) represents the number of columns of the adjacency matrix A; W (K) is the parameter matrix of the Kth graph convolutional layer, W (K) The two dimensions of the parameter matrix are consistent with the node feature dimensions; σ represents the Relu nonlinear activation function; After two layers of graph convolution were performed on the user relationship graph and the item relationship graph respectively, the user relationship information and item relationship information were fully modeled; 2). UI interaction feature extraction module After the user relationship features and item relationship features are fully modeled, for the paired target user u i and items to be recommended v j , through a layer of graph interaction layer, the relevant node relationship features are fused to obtain the corresponding interaction feature expression; first, the target user relationship feature h is concatenated i u and the relationship features of the recommended item and its first-order neighbor nodes; input the concatenated features into the attention network, and perform softmax processing on the network output value to obtain the weight coefficient of the relationship features of the recommended item and its first-order neighbor nodes; and obtain the recommended item v by weighting and summing the corresponding item relationship features. j The mathematical expression of the interaction characteristics of the part is: Among them, W1, W2 represent the parameter matrix of the two-layer attention network, b1, b2 represent the deviation coefficients of the two-layer attention network, σ represents the nonlinear activation function, and the Relu activation function is used; It represents the target user u in the user relationship graph i relationship characteristics; Represents item v in the item relationship graph t The relationship characteristics of node t(v t ) belongs to the recommended items v j Any item node in its first-order neighbor item set; N(j) represents the item v to be recommended j The set of first-order neighbors of ; Indicates splicing processing; is the output value of the attention network, and the input of the network is the target user relationship feature and itemsv t Relationship characteristics The concatenated feature vectors are then normalized by softmax on all output values ​​of the attention network to obtain the item node v t The corresponding weight coefficient μ it ; Formula (11) represents the softmax processing, and the denominator is the node v of the item to be recommended j The sum of the attention network output values ​​corresponding to all its first-order neighbor item nodes, with the numerator being the item node v t The output value of the attention network; where the denominator x represents the recommended item v j Any item node of its first-order neighbor item set; and The calculation method is the same as that of the item node x(v x ) and target user u i Relationship characteristics After splicing, it is input into the attention network to obtain the output value of the attention network; finally, according to the obtained item node v t The weight μ it , weighted sum of all relevant item nodes v t Relationship characteristics Get the recommended item v j Partial interaction feature z j ; Similarly, the relationship characteristics of the recommended items and the target user and its first-order neighbor node relationship features; input the concatenated features into the attention network, and perform softmax processing on the network output value to obtain the weight coefficient of the target user and its first-order neighbor node relationship features; obtain the target user u by weighted summing up the corresponding user relationship features i The mathematical expression of the interaction characteristics of the part is: Among them, W1, W2 represent the parameter matrix of the two-layer attention network, b1, b2 represent the deviation coefficients of the two-layer attention network, σ represents the nonlinear activation function, and the Relu activation function is used; Represents the item relationship graph to be recommended item node v j relationship characteristics; Represents user node u in the user relationship graph k The relationship characteristics of node k(u k ) belongs to the target user u i Any user node in its first-order neighbor user set; N(i) represents the target user u i The set of first-order neighbor nodes; Indicates splicing processing; is the output value of the attention network, and the input of the network is the relationship feature of the items to be recommended and user u k Relationship characteristics The concatenated feature vectors are then normalized by softmax on all output values ​​of the attention network to obtain the user node u k The corresponding weight coefficient β jk ; Formula (14) represents softmax processing, and the denominator is the target user node u i The sum of the attention network output values ​​corresponding to all its first-order neighbor user nodes, with the numerator being the user node u k The output value of the attention network; where the denominator y represents the target user u i Any user node in its first-order neighbor user set; and The calculation method is the same as that of the user node y(u y ) Relationship features with items to be recommended After splicing, it is input into the attention network to obtain the output value of the attention network; finally, according to the obtained user node u k The weight β jk , weighted sum of all relevant user nodes u k Relationship characteristics Get the target user u i Partial interaction feature z i ; After getting the target user u i Partial interactive feature expression z i 、Items to be recommendedv j Partial interactive feature expression z j Finally, the interaction features of the two parts are concatenated to obtain the target user u i and items to be recommended v j The interactive characteristics of; its mathematical expression is: z i is the interaction feature of the user part, z j is the interactive feature of the part to be item; Indicates splicing processing; z ij For the final target user u i And its recommended items v j Interactive feature expression of 3). Interactive inference module After getting the target user u i and items to be recommended v j After the interaction features are obtained, the model learns the distribution law of the interaction features through the interaction inference module, and then obtains the user u i For items v j Accurate preference prediction; Interaction inference module This method uses the classic DNN network; directly input the interaction features into the DNN network to obtain the user's predicted score for the item, and then use sigmoid normalization to model the model's predicted score as user u i For items v j The preference probability expression of is: g1=z ij (17) g2=σ(W1·g1+b1) (18) g3=σ(W2·g2+b2) (19) r ij '=sigmoid(W3·g3) (20) Among them, z ij For target user u i And its recommended items v j W1, W2, W3 represent the parameter matrix of the DNN network, b1, b2 represent the deviation coefficients in the DNN network, σ represents the nonlinear activation function, and the Relu activation function is used; g1, g2, g3 are the interactive vector expressions of each layer of the DNN network output; r ij ' is the prediction evaluation expression of the model after sigmoid normalization, where the sigmoid in formula (20) represents the normalized mapping of the output value of the last layer of the DNN network to the sigmoid function; 4). Top-N recommendation module After obtaining the model's predicted ratings for all items of each user, the items participating in the rating are sorted from large to small according to the model output values. The top-N recommendation mechanism is used to recommend the top N items after sorting to the target user, thereby completing personalized recommendations for the user.

Images