User preference prediction method based on graph neural network conversation recommendation system

Abstract

The application discloses a user preference prediction method based on a graph neural network conversation recommendation system, and relates to the technical field of artificial intelligence, comprising the following steps: S1, obtaining a training data set; S2, preprocessing the training data set to obtain a conversation sequence; S3, constructing a preference prediction model; S4, importing the conversation sequence into the preference prediction model and training and optimizing; S5, obtaining user data to be predicted, preprocessing the user data to be predicted to obtain a conversation sequence to be predicted; and S6, importing the conversation sequence to be predicted into the optimized preference prediction model to obtain a prediction result. In the process of constructing a conversation graph, a self-loop is added to each item node, the relationship of edges is obtained, the information of an adjacency matrix is increased, the matrix sparsity problem is solved, a multi-channel graph neural network is used to extract local interest features of the conversation, the information in the conversation graph is efficiently utilized, then a multi-layer self-attention network is introduced to extract global interest features of the conversation, and accurate and effective preference prediction is realized through the conversation.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a method for predicting user preferences based on a graph neural network-based conversational recommendation system. Background Technology

[0002] Recommender systems are decision support systems that predict a user's interest in specific items and help the user select content of interest based on information such as historical behavior and preferences. Traditional recommender systems require long-term historical interaction data to learn user behavior and preferences. However, in some cases, obtaining this data may be limited by user privacy issues, making traditional recommender systems ineffective. Conversational recommender systems are an important branch of recommender system research. They do not rely on long-term historical interaction data; they only need to generate recommendations for users based on short-term interaction information of the current user.

[0003] Common research methods can be divided into two main categories: conversation recommendation algorithms based on traditional methods and conversation recommendation algorithms based on deep learning.

[0004] Traditional methods primarily utilize data mining or machine learning techniques to capture embedded dependencies in session data. These methods can be broadly categorized into three types: collaborative filtering-based methods, K-nearest neighbor-based methods, and Markov chain-based methods. Sarwa et al. proposed an item-based collaborative filtering algorithm that effectively addresses online computational performance issues. Amazon designed an item-based collaborative filtering recommendation system based on this, primarily recommending products based on user behavior. Liang et al. proposed a matrix factorization-based collaborative filtering algorithm, combining matrix factorization techniques with collaborative filtering to simultaneously capture the user's clicked item set and user interest transitions. Rendle et al., in their session recommendation system, introduced a personalized transition matrix using Markov chains to simultaneously capture temporal information and long-term user preference information, while also incorporating a matrix factorization model to effectively address the sparsity problem of the transition matrix.

[0005] With the development of deep learning technology, more and more scientific fields are using deep learning-based methods to solve academic challenges.

[0006] Graph Neural Networks (GNNs) differ from traditional deep learning neural networks such as convolutional neural networks and recurrent neural networks in that they can effectively model non-Euclidean spatial data, learning feature information from the neighboring nodes of the target node. Wu et al. first applied GNNs to the conversation recommendation task, proposing a conversation recommendation algorithm based on gated graph neural networks (SR-GNN). This algorithm was the first to apply GNNs to conversation recommendation tasks, marking a pioneering and milestone achievement. Subsequently, numerous scholars have conducted research on conversation recommendation systems based on GNNs. Xu et al. proposed the GC-SAN model, which, compared to SR-GNN, uses multi-layer self-attention to capture features at different levels from modules at different heights. Lu et al., drawing on some of the structure of the GC-SAN model, proposed the CIE-GNN model, which uses multi-head attention and additive attention to learn the global representation vector of the conversation. Yu et al. proposed the TAGNN model, employing a gated graph neural network, and further improved the conversation representation by proposing an attention network model for the target item to adapt to the constantly changing intent of the user during the conversation.

[0007] Existing graph neural network-based conversation recommendation algorithms suffer from low information utilization in the conversation graph during construction, resulting in limited information available for item conversion relationship modeling. Furthermore, the simplistic graph construction method leads to sparse adjacency matrices in the stored conversation graph. These issues ultimately result in poor recommendation performance. Summary of the Invention

[0008] The purpose of this invention is to design a user preference prediction method based on a graph neural network-based conversational recommendation system in order to solve the above-mentioned problems.

[0009] The present invention achieves the above objectives through the following technical solutions:

[0010] User preference prediction methods based on graph neural network-based conversational recommendation systems include:

[0011] S1. Obtain the training dataset;

[0012] S2. Preprocess the training dataset to obtain the session sequence;

[0013] S3. Construct a preference prediction model. The preference prediction model includes a conversation graph construction layer, a local encoding layer, a global encoding layer, and a prediction output layer from input to output. The local encoding layer is a multi-channel conversation graph neural network, and the global encoding layer is a multi-layer attention neural network. The local encoding layer is used to encode the local interests of the conversation graph to obtain the short-term local interest features of the conversation graph. The global encoding layer encodes the global interest preferences of the conversation graph to obtain the global interest features of the conversation graph. The prediction layer obtains the prediction result based on the short-term local interest features and the global interest features.

[0014] S4. Import the session sequences from the training set into the preference prediction model, and use the cross-entropy function to train and optimize the preference prediction model.

[0015] S5. Obtain the user data to be predicted and preprocess it to obtain the session sequence to be predicted;

[0016] S6. The conversation sequence to be predicted is imported into the optimized preference prediction model to obtain the prediction results.

[0017] The beneficial effects of this invention are as follows: By focusing on the construction of the conversation graph, the encoding of local interest in the conversation, and the encoding of global interest in the conversation, a self-loop is added to each item node when constructing the conversation graph, and the relationship of the edges is divided into four categories: outgoing edges, incoming edges, outgoing-incoming edges, and self-loops. This effectively increases the information of the adjacency matrix and solves the problem of matrix sparsity. A multi-channel graph neural network is used to extract the local interest features of the conversation, which makes more efficient use of the information in the conversation graph. A multi-layer self-attention network is introduced to extract the global interest features of the conversation, which can learn the potential different types of attention points in the sequence and adaptively assign weights to each item in the conversation, thus realizing accurate and effective preference prediction through the conversation. Attached Figure Description

[0018] Figure 1 This is a flowchart of the user preference prediction method based on a graph neural network-based conversational recommendation system according to the present invention;

[0019] Figure 2 This is a schematic diagram of the session graph construction of the present invention;

[0020] Figure 3 This is the flowchart of the first model in the experiment;

[0021] Figure 4 This is the flowchart of the second model in the experiment. Detailed Implementation

[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0023] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0024] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.

[0025] In the description of this invention, it should be understood that the terms "upper," "lower," "inner," "outer," "left," "right," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of this invention is in use, or the orientation or positional relationship commonly understood by those skilled in the art. They are only used to facilitate the description of this invention and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0026] Furthermore, the terms "first," "second," etc., are used only to distinguish descriptions and should not be interpreted as indicating or implying relative importance.

[0027] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, terms such as "set" and "connection" should be interpreted broadly. For example, "connection" can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection or an indirect connection through an intermediate medium; it can be a connection within two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0028] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0029] like Figure 1 As shown, the user preference prediction method based on graph neural network-based conversational recommendation system includes:

[0030] S1. Obtain the training dataset;

[0031] S2. Preprocess the training dataset to obtain the session sequences; specifically including:

[0032] S21. Reconstruct the item IDs starting from 1;

[0033] S22. Filter out items with a cumulative number of clicks less than 5 and sessions with a number of clicked items less than 2 in all sessions. This effectively prevents the model from failing to learn the transformation relationship between items in the session and the feature information of the session due to the session sequence being too short.

[0034] S23. Perform data augmentation on the filtered sessions to expand the dataset, making a session v i'=(v i,1 ,v i,2 ,v i,3 ,…,v i,t ) transformed into multiple sub-sessions ({v i,1},v i,2 ), ({v i,1 ,v i,2},v i,3 ), ..., ({v i,1 ,v i,2 ,v i,3 ,…,v i,t-1},v i,t ); where () contains two elements, where the element in {} represents the training data, and the other represents the data label corresponding to the training set, which together form the conversation sequence s = {v1, v2, v3, ..., v t}, where t represents the number of sessions, v i =({v i,1 ,v i,2 ,v i,3 ,…,v i,i-1},v i,i ).

[0035] S3. Construct a preference prediction model. The preference prediction model includes a conversation graph construction layer, a local encoding layer, a global encoding layer, and a prediction output layer from input to output. The local encoding layer is a multi-channel conversation graph neural network, and the global encoding layer is a multi-layer attention neural network. The local encoding layer is used to encode the local interests of the conversation graph to obtain the short-term local interest features of the conversation graph. The global encoding layer encodes the global interest preferences of the conversation graph to obtain the global interest features of the conversation graph. The prediction layer obtains the prediction result based on the short-term local interest features and the global interest features.

[0036] S4. Import the session sequences from the training set into the preference prediction model, and train and optimize the preference prediction model using the cross-entropy function. The cross-entropy loss function is expressed as: Among them, y i It is the actual next clicked item tag corresponding to the session, in the form of a one-hot encoded vector. This refers to the probability value predicted by the preference prediction model, where n is the total number of items in the item set V, and J is the loss value; specifically, it includes:

[0037] S41. The session graph construction layer constructs a session sequence s = {v1, v2, v3, ..., v...} for a given session. t}, the item v in this sequence i Arranged according to the order in which users clicked, as follows: Figure 2As shown, a session graph G is constructed for the sequence s = {v1, v3, v4, v2, v3, v1}. s =(V s E s Vs = {v1, v2, v3, v4} is the set of nodes in the conversation graph, and Es = E s1 ∪E s2 It is the set of edges in the conversation graph, E s1 ={(v1,v3),(v3,v4),(v4,v2),(v2,v3),(v3,v1)} is the set of directed edges formed by items in the conversation sequence according to their chronological order. s2 ={(v1,v1),(v2,v2),(v3,v3),(v4,v4)} is a self-loop for each item, and each session corresponds to an adjacency matrix A. s The matrix contains five types of values. A[i][j] = 0 indicates that there is no edge between node i and node j. A[i][j] ∈ {1,2,3,4} corresponds to four types of edges in the conversation graph, namely self-loop, incoming edge, outgoing edge, and in-outgoing edge.

[0038] S42. Analyze the session set S = {s1, s2, s3, ..., s} m The items contained in} constitute an item set item = {i1, i2, i3, ..., i k}, where k represents the total number of k distinct item items; each of the k item items is randomly initialized with an embedding vector, forming a randomly initialized embedding representation matrix E = Embedding(num,d), where E∈R num＊d num is the number of different types of items in the dataset, d is the dimension of the embedding vector, and the session sequence s = {v1, v2, v3, ..., v t The corresponding embedding representation matrix is ​​represented as X = E[v1, v2, v3, ..., v t ], where v i It is the label of the i-th element in sequence s, and this label corresponds to the row of matrix E;

[0039] S43. The local coding layer encodes the local information of session s to obtain the local coding result. Represented as Among them, a ij Represents node v j The feature vector for node v i In terms of its weight value, It is node v in session s i The set of neighboring nodes; the feature encoding of the last item in session s is selected as the short-term local interest feature of the session, denoted as... Specifically:

[0040] ① Node v i Embedding vector and node v j The embedding vector is multiplied by the vector to obtain a vector of shape (d, 1). This vector is then subjected to a linear transformation and a nonlinear transformation to obtain node v. i The weight coefficient e of the feature vector for the node ij , represented as Where, r ij Represents node v i With node v j The relationship between edges, where each type of edge corresponds to a parameter vector a∈R. d There are a total of four types of parameter vectors: a self ,a in ,a out ,a in-out These correspond to four different channels;

[0041] ② Use the Softmax function to adjust the weight coefficients e ij Normalization is performed to obtain node v j The feature vector for node v i In terms of its weight value a ij , represented as Make the weight value e ij It is comparable across different nodes;

[0042] ③ Aggregation node v i All neighboring nodes, each neighboring node has a relationship with node v. i The weights of the nodes in the current session s are summed over all neighboring nodes to obtain the value of node v. i Item coding results

[0043] ④ Select the feature encoding of the last item in session s as the short-term local interest feature of the session, denoted as:

[0044] ⑥ The local coding layer encodes local information for session s, represented as: Among them, e ij Represents node v j The feature vector for node v i To describe its weight value, each type of edge corresponds to a parameter vector a∈R. d There are a total of four types of parameter vectors: a self ,a in ,a out ,a in-outThese correspond to four different channels; It is node v in session s i The set of neighboring nodes; the feature encoding of the last item in session s is selected as the short-term local interest feature of the session, denoted as...

[0045] S44. The global encoding layer uses a self-attention mechanism to encode the item feature matrix H = [h1, h2, h3, ..., h...] of the session set S. t ], H∈R d＊n The global encoding result E is obtained by performing self-attention calculation, denoted as E = MLSA(H), E k =MLSA(E k-1 ), where E k ∈R n＊d E1 is the global encoding vector of the session generated through a K-layer attention mechanism, where E1 = E; k The feature vector corresponding to the last item in the session is taken as the global interest feature of the session, and can be represented as: Specifically:

[0046] 1) For H = [h1, h2, h3, ..., h t ], H∈R d＊n The output matrix F obtained by performing self-attention mechanism processing once is represented as:

[0047] Q = W Q H

[0048] K = W K H

[0049] V = W V H

[0050]

[0051]

[0052] Among them, W Q W K W V ∈R 2d＊d Here are learnable parameters, Q is the query matrix, K is the keyword matrix, V is the value matrix, and A∈R. n＊n It is a normalized attention weight matrix that contains the importance of all other nodes in the session to the current node;

[0053] 2) Two linear transformations and one nonlinear transformation are added to the output matrix F, and then the original matrix information is added. Simultaneously, borrowing from residual networks, the item embedding matrix H of the session is obtained and given a layer of self-attention global encoding result E1, expressed as E1 = dropout(ReLU(W1F+b1)W2+b2)+F). This enhances the fitting ability of the preference prediction model, where ReLU is the activation function, and W1, W2, b1, and b2 are learnable parameters, W1, W2 ∈ R. 2d*d b1, b2∈R d ;

[0054] 3) The global encoding vector E of the session generated by the K-layer attention mechanism k E k =MLSA(E k-1 ), E1 = E, and features at different levels have been learned;

[0055] 4) E k The feature vector corresponding to the last item in the session is taken as the global interest feature of the session, and can be represented as:

[0056] S45. The prediction layer combines short-term local interest features and global interest features into a vector s by weighted summation. final , represented as s final =ω*s local +(1-ω)*s global This vector represents the final interest preferences of the session. final s final The final interest preference vector of the session is multiplied by the embedding matrix E of the item set V, and then the recommendation probability of the session for each item in the item set V is obtained by applying the softmax function. Represented as Where ω∈(0,1) is an adjustable hyperparameter, and E is the embedding matrix corresponding to the item set V.

[0057] S5. Obtain the user data to be predicted and preprocess it to obtain the session sequence to be predicted;

[0058] S6. The conversation sequence to be predicted is imported into the optimized preference prediction model to obtain the prediction results.

[0059] Verification Experiment

[0060] To validate the effectiveness of the preference prediction model, its performance was tested on two public standard datasets. The task of a conversational recommendation algorithm is to predict the item a user will click next, based on the conversation generated by the current user's interaction with the system. Typically, more than one recommendation item is returned to the user; to include as many items as possible of the user's interest, a recommendation list is generated and returned. For this, two common evaluation metrics are MRR@N and P@N. MRR@N (Mean Reciprocal Rank) indicates that each sample label ranks highly in the generated recommendation list in the test set, indicating more accurate recommendations. P@N (Precision@N) refers to the accuracy of the top N recommended items in the recommendation list; a higher value indicates better model recommendation performance.

[0061] Table 1 shows the test results of our proposed preference prediction model and several representative methods in conversational recommendation algorithms on two standard public datasets. As can be seen from the table, our proposed method achieves the best results across all metrics. Compared with the representative graph neural network-based conversational recommendation algorithm SR-GNN, our model comprehensively outperforms the SR-GNN model in all four metrics: P@10, MRR@10, P@20, and MRR@20. For the Diginetica dataset, the improvements are 4.2%, 2.52%, 3.43%, and 1.39%, respectively; and for the Nowplaying dataset, the improvements are 2.88%, 0.24%, 4.26%, and 0.31%, respectively, demonstrating the effectiveness of our proposed preference prediction model.

[0062] To investigate the effectiveness of the information propagation module in graph neural networks, two additional comparative models were constructed. The first model, without using a multi-channel module (i.e., without utilizing network information from the session graph), directly feeds the session embedding matrix into a multi-layer self-attention mechanism for global interest and preference extraction. This model is referred to as M in this paper. only-att The second model, in this method, replaces the multi-channel graph neural network in the MCSA-GNN-SR model with a gated graph neural network from the SR-GNN model. Simultaneously, the adjacency matrices of the constructed session graph and the storage graph are also modified to make them suitable for the gated graph neural network. This model is referred to as M in this paper. GGNN-att The specific calculation process for the two models is as follows: Figure 3 and Figure 4As shown in Table 2, experiments were conducted on the two comparative models mentioned above on the standard datasets Diginetica and Nowplaying. Four evaluation metrics—P@10, MRR@10, P@20, and MRR@20—were used to compare them with the MCSA-GNN-SR model presented in this chapter. The specific experimental results are shown in Table 2. The MCSA-GNN-SR model outperforms both comparative models on the datasets, comprehensively demonstrating the effectiveness of the graph neural network information propagation module designed in the MCSA-GNN-SR model of this chapter.

[0063] Table 1 Comparison of experimental results of MCSA-GNN-SR model

[0064]

[0065] Table 2M o晦ly-att M GGNN-att Comparison of MCSA-GNN-SR three-model experiments

[0066]

[0067] The technical solutions of the present invention are not limited to the specific embodiments described above. Any technical modifications made in accordance with the technical solutions of the present invention fall within the protection scope of the present invention.

Claims

1. A user preference prediction method based on a graph neural network-based conversational recommendation system, characterized in that, include: S1. Obtain the training dataset; S2. Preprocess the training dataset to obtain the session sequence; S3. Construct a preference prediction model. The preference prediction model includes a conversation graph construction layer, a local encoding layer, a global encoding layer, and a prediction output layer from input to output. The local encoding layer is a multi-channel conversation graph neural network, and the global encoding layer is a multi-layer attention neural network. The local encoding layer is used to encode the local interests of the conversation graph to obtain the short-term local interest features of the conversation graph. The global encoding layer encodes the global interest preferences of the conversation graph to obtain the global interest features of the conversation graph. The prediction layer obtains the prediction result based on the short-term local interest features and the global interest features. S4. Import the session sequences from the training set into the preference prediction model, and use the cross-entropy function to train and optimize the preference prediction model; specifically including: S41. Session graph construction layer for a session construction sequence Items in this sequence Arrange the sequences according to the order in which users clicked, and construct a session graph for the sequence s. V s It is the set of nodes in the session graph, E s It is the set of edges in the session graph, where each session corresponds to an adjacency matrix A. s The matrix contains five types of values. This indicates that there is no edge between node i and node j. These correspond to the four types of edges in the conversation graph: self-loop, incoming edge, outgoing edge, and in-outgoing edge. S42, Analyze the session set The items contained herein constitute an item set item={ } where k represents the total number of k distinct items; each of the k items is randomly initialized with an embedding vector, forming a randomly initialized embedding representation matrix. , where E∈ num is the number of different types of items in the dataset, d is the dimension of the embedding vector, and the embedding representation matrix corresponding to the session sequence s is represented as follows: ,in It is the label of the i-th element in sequence s, and this label corresponds to the row of matrix E; S43. The local coding layer encodes local information for session s, obtaining the local coding result. , represented as ,in, Represents a node The feature vector for the node In terms of its weight value, It is a node in session s The set of neighboring nodes; the feature encoding of the last item in session s is selected as the short-term local interest feature of the session, denoted as... ; S44. The global encoding layer uses a self-attention mechanism to encode the item feature matrix of the session set S. , ∈ The global encoding result E is obtained by performing self-attention calculation, denoted as: , ,in, E1 is the global encoding vector of the session generated through a K-layer attention mechanism, where E1 = E; The feature vector corresponding to the last item in the session is taken as the global interest feature of the session, and can be represented as: ; S45. The prediction layer combines short-term local interest features and global interest features into a vector s by weighted summation. final , represented as This vector represents the final interest preferences of the session. final s final The final interest preference vector of the session is multiplied by the embedding matrix E of the item set V, and then the recommendation probability of the session for each item in the item set V is obtained by applying the softmax function. , represented as ;in, ∈(0,1) is an adjustable hyperparameter, and E is the embedding matrix corresponding to the item set V; S5. Obtain the user data to be predicted and preprocess it to obtain the session sequence to be predicted; S6. Import the conversation sequence to be predicted into the optimized preference prediction model to obtain the prediction results.

2. The user preference prediction method based on a graph neural network-based session recommendation system according to claim 1, characterized in that, S2 includes: S21. Reconstruct the item IDs starting from 1; S22. Filter out items with a cumulative number of clicks less than 5 and sessions with a number of clicked items less than 2 from all sessions; S23. Perform data augmentation on the filtered sessions, transforming a session Transformed into multiple sub-sessions ({ }, ), ( ) 、...、( ); where there are two elements in parentheses, the element in curly braces {} represents the training data, and the other represents the data label corresponding to the training set, which are combined to form the conversation sequence. Where t represents the number of sessions, =( ).

3. The user preference prediction method based on a graph neural network-based session recommendation system according to claim 1, characterized in that, In S4, the cross-entropy loss function is used to optimize the preference prediction model. The cross-entropy loss function is expressed as: ,in, It is the actual next clicked item tag corresponding to the session, in the form of a one-hot encoded vector. This is the probability value predicted by the preference prediction model, where n is the total number of items in the item set V. It is the loss value.

4. The user preference prediction method based on a graph neural network-based session recommendation system according to claim 1, characterized in that, S43 includes: ① Node Embedded vectors and nodes The embedding vector is multiplied by a dot product to obtain a vector of shape (d, 1). This vector is then subjected to a linear transformation and a nonlinear transformation to obtain the nodes. The weight coefficient e of the feature vector for the node ij , represented as , where r ij Represents a node With nodes The relationship between edges, each type of edge corresponds to a parameter vector. There are a total of four types of parameter vectors: These correspond to four different channels; ② Use the Softmax function to adjust the weight coefficients e ij Normalization is performed to obtain nodes The feature vector for the node In terms of its weight value a ij , represented as ; ③ Aggregation node v i All neighboring nodes, each neighboring node has a relationship with node v. i The weights of the nodes in the current session s are summed over all neighboring nodes to obtain the value of node v. i Item coding results , ; ④ Select the feature encoding of the last item in session s as the short-term local interest feature of the session, denoted as: ; ⑤ The local coding layer encodes local information for session s, represented as: , where e ij Represents a node The feature vector for the node To describe its weight value, each type of edge corresponds to a parameter vector. There are a total of four types of parameter vectors: These correspond to four different channels; It is a node in session s The set of neighboring nodes; the feature encoding of the last item in session s is selected as the short-term local interest feature of the session, denoted as... .

5. The user preference prediction method based on a graph neural network-based session recommendation system according to claim 1, characterized in that, S44 includes: 1) To , ∈ The output matrix F obtained by performing self-attention mechanism processing once is represented as: ； ； ； ； ； in, ∈ These are learnable parameters, where Q is the query matrix, K is the keyword matrix, V is the value matrix, and A∈ It is a normalized attention weight matrix that contains the importance of all other nodes in the session to the current node; 2) Add two linear transformations and one nonlinear transformation to the output matrix F, and then add the original matrix information to obtain the item embedding matrix H of the session. Then apply a layer of self-attention to the global encoding result E1, denoted as... ; 3) The global encoding vector of the session generated by the K-layer attention mechanism k , E1 = E; 4) The feature vector corresponding to the last item in the session is taken as the global interest feature of the session, and can be represented as: .

Images