A conversation recommendation method based on a sequence growth perception model

Abstract

The application relates to a conversation recommendation method based on a sequence growth perception model and belongs to the technical field of network big data information recommendation. In the method, all conversations are divided into multiple growth states according to a window increment rule with a size of one. The same growth state of all conversations is established as a global dynamic graph in a discrete form to model sequential information and dynamic transition patterns. Then, a dynamic graph neural network containing a growth perception layer and a dynamic perception layer is used to capture the global dynamic transition patterns. Meanwhile, a single conversation is established as a local static graph to capture local static transition patterns. Finally, conversation recommendation is realized based on the captured global dynamic transition patterns and local static transition patterns. The method greatly enriches the item transition pattern information in the conversation and effectively improves the recommendation effect.

Description

TECHNICAL FIELD

[0001] The application relates to a session recommendation method based on a sequence growth perception model and belongs to the technical field of network big data information recommendation. BACKGROUND

[0002] With the development of the Internet and mobile devices, a large amount of data is generated in the network. When users face choices, these data cause serious information overload. A recommendation system can recommend the most interesting and most relevant data information to users according to relevant information, effectively alleviating the information overload problem, and is therefore widely used in e-commerce, search engines, video recommendation and other application scenarios.

[0003] In order to achieve efficient and accurate recommendation, at present, many representative methods have been proposed, which can be divided into three types of collaborative filtering, content-based recommendation and hybrid recommendation. Among them, collaborative filtering refers to a recommendation system that scores or recommends unknown items according to existing scores; the content-based recommendation method considers additional user configuration information and item attribute information to generate recommended items; hybrid recommendation is a combination of the two to enhance the recommendation effect.

[0004] Although the above methods have been applied in the academic and industrial fields, there are still some deficiencies. These methods assume that historical interactions and recent interactions are equally important, but ignore the fact that user preferences are dynamic and change over time, that is, the importance of recent interactions is obviously stronger than long-term historical interactions. In addition, due to privacy policies and service type restrictions, user profiles and long-term historical interactions cannot be obtained in some cases, resulting in a serious decline in the performance of these methods.

[0005] Due to the above limitations, recommendation using anonymous short sequence recent interaction sequences, that is, session-based recommendation methods, have received more and more attention in recent years. Sessions are obtained by naturally dividing recent interaction records of anonymous users, and the symbiotic and sequential information between interaction items is well preserved. Based on these session data, the recommendation system can capture the complex item transition relationship and accurately model the user's recent preferences and trends. The session-based recommendation method can capture the trend of user preferences based on simple data, achieve fast and accurate interaction recommendation, and has very important significance for academic research and industrial application.

[0006] The existing session-based recommendation methods can be divided into three types according to their technical characteristics: traditional methods, deep learning-based methods and graph neural network-based methods. The traditional method is relatively simple, easy to understand and implement, and is very effective on the dataset with simple dependency, but lacks the understanding ability of complex conversations; the method based on deep learning technology can model the complex transfer mode within and between sessions, but usually requires a large amount of calculation and complex network design, which limits the model expression ability.

[0007] The method based on graph neural network can model the non-sequential transfer mode and high-order dependency between items, but since it still regards a complete session as a static item transfer mode, it ignores the dynamic characteristics of the sequence growing with interactive operations, and has the following problems:

[0008] (1) Loss of sequential information. The establishment of the conversation graph completely depends on the researchers' cognition and assumption of the item association relationship in the session sequence, so under the premise of regarding the session as a complete static sequence, the graph structure established is always the static structure corresponding to the complete session, but the static graph structure may ignore some sequential information existing in the session sequence.

[0009] (2) Loss of dynamic transfer mode. Each growth of the session sequence can reflect the influence of the existing interaction on the new interaction, and also reflect the subsequent influence of the growing sequence on the next interaction, and the key to modeling this dynamic transfer mode is that after the new item is added, the new item must update itself based on the existing items, and the existing items also update themselves according to the newly added items. However, in the graph structure constructed by the method ignoring the sequence growth characteristics, the association relationship of all nodes is determined before the information propagation and aggregation operation, and the information update of the corresponding nodes after the new interaction is added cannot be completed. SUMMARY

[0010] The purpose of the present application is to solve the technical problems of sequential information loss and dynamic transfer mode loss caused by ignoring the sequence growth characteristics in network session recommendation, and innovatively proposes a session recommendation method based on a sequence growth perception model, which can significantly improve the accuracy of session recommendation.

[0011] The core idea of the application is: all sessions are divided into multiple growth states according to the window increment rule with a size of one. The same growth state of all sessions is established as a discrete global dynamic graph to model the sequential information and dynamic transition mode (each session is divided according to the established rule, and the incomplete session obtained by division is called a growth state. Therefore, multiple corresponding incomplete sessions are constructed into a graph structure, and the graph structure is a dynamic graph with a discrete form). Then, the dynamic graph neural network containing the growth perception layer and the dynamic perception layer is used to capture the global dynamic transition mode. At the same time, a single session is established as a local static graph for capturing the local static transition mode. Finally, based on the captured global dynamic transition mode and local static transition mode, the session recommendation is realized.

[0012] The application is realized based on the following technical solutions.

[0013] A session recommendation method based on a sequence growth perception model, comprising the following steps:

[0014] Step 1: constructing a global dynamic graph using a session.

[0015] All sessions are divided according to the established rule, and the growth state obtained by division is established as a discrete dynamic graph structure, which is called a global dynamic graph.

[0016] All sessions refer to all session sequences in the data set. The purpose of using all sessions is to introduce the cross-session item transition mode, that is, a global level information.

[0017] The established rule for dividing the session is: a division window with an initial size of two and a gradient of one size is increased, which can divide each session into multiple growth states.

[0018] The discrete dynamic graph is a specific dynamic graph representation form, and the dynamic nature of the graph structure is embodied by multiple sequentially arranged nodes and edges of different static graphs. Through the evolution of this graph structure, the growth process of the sequence is explicitly modeled, solving the modeling problem of sequential information and dynamic transition mode.

[0019] A static graph is composed of the same growth state of several sessions, and each node in it represents an item. The adjacent edge between two nodes represents the existence of adjacent interaction order between two items, and the number of edges between two nodes represents the number of adjacent appearances of two items.

[0020] Step 2: constructing a local static graph using a single session.

[0021] A single session refers to an independent session sequence and does not contain interactions with other sessions, that is, a local level information.

[0022] The established static graph is a kind of undirected graph structure, in which the nodes correspond to items, the edges correspond to the adjacent relationship between two items, and the number of edges is consistent with the number of times of adjacent occurrence of two items.

[0023] Step 3: Structure growth perception of the global dynamic graph is performed.

[0024] In order to capture the dynamic transfer pattern of items in the process of multiple graph structure changes, the structure information of each static graph in the global dynamic graph established in step 1 is accurately extracted, and the extracted structure information is subjected to corresponding position embedding.

[0025] For a static graph structure, first, the node vector matrix is obtained through the embedding layer, and then the information propagation and aggregation operation between nodes is completed according to the adjacency matrix, so as to update the vector matrix of the node. The final vector matrix contains the corresponding graph structure information.

[0026] When performing the information propagation and aggregation operation, the degree of association between nodes, i.e. the association coefficient of the node and its neighbor, is considered, and then the weighted sum is performed according to the association coefficient to complete the targeted information propagation and aggregation.

[0027] The above structure information extraction operation is repeated to obtain the node vector matrix corresponding to the number of static graphs.

[0028] Although there is no order information in the nodes inside the static graph, there is an order corresponding to the sequence growth between multiple static graphs. Therefore, the corresponding position vector is embedded on the basis of the node vector matrix to obtain the final growth perception output.

[0029] Step 4: Structure dynamic perception of the global dynamic graph is performed.

[0030] After obtaining the node vector matrix of each static graph containing structure information and order information, in order to capture the overall structure dynamic change, the self-attention mechanism is used to associate and fuse multiple node vector matrices.

[0031] Since the number of nodes in multiple static graphs meets the incremental rule, i.e. the nodes in the subsequent graph must contain all the nodes in the previous graph, the shapes of the finally obtained multiple node vector matrices are inconsistent. In order to facilitate structure dynamic perception, other matrices are padded to the same size according to the number of nodes in the last node vector matrix. The vector of the padded node can be a zero vector.

[0032] For the same node, the information obtained in different graphs is different, so the corresponding vector of the same node in multiple vector matrices contains multiple different structural information, and the use of self-attention to associate multiple vectors of a node can perceive the dynamic change of the structure, which represents the dynamic transfer pattern of the items in the structure evolution process.

[0033] Step 5: Capture the static transfer pattern in the local static graph.

[0034] Although the global dynamic graph obtains the global level item dynamic transfer pattern after structure growth and structure dynamic perception, it is still necessary to extract the local level static transfer pattern in a single session, and the fusion of the two can obtain a more accurate final session representation.

[0035] The global dynamic graph uses all session sequences in the construction process, which successfully expands the incremental of a single session to the overall state change of global multi-session, mines the cross-session item association, and greatly enriches the dynamic transfer pattern contained in the item node. However, for a session-based recommendation system, the task is to give the next interaction object of a single session, and the global information across sessions may contain a lot of noise information irrelevant to the session, so it is still necessary to learn the item transfer pattern in a single session as local static information and global dynamic information fusion, and finally obtain a balanced information session representation for the next interaction prediction.

[0036] For a single session, it is more efficient to build it into a graph structure and use a graph neural network to extract the transfer pattern. However, in the specific extraction process, since the main purpose of extracting the local static transfer pattern is to obtain a more accurate session representation, the degree of association between the node and the overall session representation also needs to be considered.

[0037] Therefore, similarity calculation is performed according to the node vector and the session average representation to obtain a pair of association coefficients, so that nodes similar to the session will be assigned a larger attention weight.

[0038] Step 6: Session representation based on global dynamic transfer pattern and local static transfer pattern.

[0039] From the multiple node vector matrices obtained in step 4, the node vector matrix corresponding to the last graph structure is selected as the global dynamic transfer pattern. Step 5 obtains the local static transfer pattern from a single session. The two are fused through a fully connected layer to obtain the final session representation. The session representation considers both global and local level information and dynamic and static transfer patterns.

[0040] Step 7: Interaction prediction using session representation.

[0041] The session representation obtained in step 6 is multiplied with the vector matrix composed of all items, and then processed by using a softmax function to obtain the probability distribution of all items becoming the next interaction. The top K items with the largest probability values are selected as the recommendation list. The value of K can be set according to the accuracy requirement.

[0042] Step 8: Model training is performed by using a cross-entropy loss function.

[0043] The cross-entropy between the real probability distribution and the predicted probability distribution obtained in step 7 is used as the loss function for model training.

[0044] By steps 1 to 8, the session recommendation based on the sequence growth perception model is completed.

[0045] Advantages

[0046] Compared with the existing session recommendation method, the present application has the following advantages:

[0047] 1. The present method selects to fuse the information of two levels of global and local, and considers to introduce two types of dynamic and static transition modes, which greatly enriches the item transition mode information in the session under the premise of ensuring information balance and accuracy.

[0048] 2. The present method proposes a discrete global session dynamic graph construction method, which divides each session into multiple growth states according to the one-size window incremental rule, and then builds a graph for the same growth state of multiple sessions. Such graph construction method not only solves the problem of loss of sequence information, but also explicitly models the growth process of the sequence, solving the problem of loss of dynamic transition mode.

[0049] 3. The present method designs a dynamic graph neural network to capture the global structure evolution dynamic information, in which the growth perception layer completes the capture of the transition mode of each graph, and the dynamic perception layer utilizes self-attention to complete the fusion of multiple modes to obtain the global dynamic transition mode. By fully mining the global level item dynamic transition mode, the recommendation effect is effectively improved, which proves the effectiveness of the item dynamic transition mode. BRIEF DESCRIPTION OF DRAWINGS

[0050] Figure 1 It is the implementation framework diagram of the method of the present application.

[0051] Figure 2 It is the global dynamic graph construction schematic diagram of the method of the present application.

[0052] Figure 3 It is the local static graph construction schematic diagram of the method of the present application. DETAILED DESCRIPTION

[0053] The method of the present application is further described in detail below in conjunction with the accompanying drawings and examples.

[0054] Examples

[0055] This example shows the specific implementation process in the real session recommendation business scenario. All the process demonstrations are based on two real high-quality data sets: Yoochoose and MovieLens-10M. The Yoochoose data set is disclosed in the $Recsys'15$ challenge, which contains six months of user click data of an online retail website. The MovieLens-10M data set is provided by the MovieLens website, which contains 10 million user movie rating data, each data consisting of user ID, movie ID, movie category, rating and timestamp.

[0056] For the Yoochoose data set, the task of session recommendation is to predict the next item that the user will click in the session according to the known item information in the session; for the MovieLens-10M data set, first, the rating data of each user is divided into multiple sessions at an interval of ten minutes, and then the next movie that the user will rate in the session is predicted according to the known movie information in the session;

[0057] Figure 1 The implementation framework diagram of the present method.

[0058] A session recommendation method based on a sequence growth perception model, in the present embodiment, comprising the following steps:

[0059] Step 1: Preprocess the two data sets.

[0060] For the two data sets, filter out the items appearing less than 5 times and the sessions with a length less than 2, respectively, and then divide the two data sets into training set and test set, respectively. At the same time, delete the items that do not appear in the training set in the test set.

[0061] Step 2: Divide the session sequence to generate data and labels for training.

[0062] For a session s, s = {v s,1 ,v s,2 ,…,v s,n}, it is divided into n data and label pairs, for example: s = {v s,1 ,v s,2 ,…,v s,n-1} is the input sequence, v s,n the next interaction item is the real label.

[0063] Step 3: Establish a global dynamic graph, such as Figure 2As shown.

[0064] Global dynamic graph G g ={g1,g2,…,g Z It consists of Z static graphs arranged in sequence, where Z is a hyperparameter representing the set number of static graphs. The process includes two steps: partitioning the session sequence and constructing the graphs.

[0065] Step 3.1: Divide the session sequence.

[0066] To explicitly model the growth of the sequence, the session is divided into multiple intermediate state sequences according to the growth order. After setting the number of static graphs Z, Z segmentation windows O with increasing lengths and a difference of 1 are obtained, O = {o1, o2, ..., o...}. Z The initial window o1 has a default length of 2. For a session s, it is divided according to the set window size to obtain a set of intermediate states m. s m s ={m1,m2,…,m Z}

[0067] Step 3.2: Construct a global dynamic graph.

[0068] After all sessions are partitioned, the intermediate states m corresponding to all sessions are... i Together they form an undirected graph g i The final global dynamic graph G is obtained. g G g ={g1,g2,…,g Z}

[0069] An undirected graph is denoted as g. i ={V i E i};V i The representative node set consists of the set m of all session sequence intermediate states. i It consists of all the items included; E i Represents the set of edges, an edge (v i,s ,v i,e () indicates that v exists in the intermediate state sequence of a certain session. i,s and v i,e Disordered pairs of items.

[0070] Step 4: Perform growth awareness of the global dynamic graph.

[0071] For the growth-aware layer, the input is the global dynamic graph G. g A static graph g i ={V i E i} and a node feature matrix d represents dimension, n represents graph gi The number of items, the vector corresponding to the item v is denoted by x v , is a vector representation.

[0072] Step 5: The importance of the relationship between one node and multiple neighbors is different, so attention mechanism is needed to calculate the weight coefficient between nodes:

[0073]

[0074] where represents the concatenation operation, is a shared parameter matrix for implementing a feedforward neural network, A uv denotes the edge weight between node u and node v in graph G i , and LeakyRelu is an activation function. At the same time, because a node can have multiple neighbors, in order to measure the importance of a neighbor node in all neighbors, the weight coefficient between nodes needs to be normalized:

[0075]

[0076] where, is the set of neighboring nodes of node v.

[0077] Step 6: After obtaining the set of neighboring nodes of node v and the weight coefficient, the neighbor node features are linearly combined according to the weight to obtain an overall representation n v of the neighbor feature, and the node itself is fused through a feedforward neural network to obtain the final node representation x′ v :

[0078]

[0079] x′ v = ReLU (W2 [x v ‖n v ]) (4)

[0080] where, a uv is the weight coefficient between node v and neighbor u, x u represents the feature vector of neighbor u, ReLU is an activation function, and x′ v is the final node representation.

[0081] Step 7: Repeat the above steps until a new node feature matrix of static graph G i is obtained

[0082] It should be noted that although there is no sequence information in the static graph, there is sequence information between the multiple static graphs constituting the dynamic graph due to the explicit incremental modeling, and the position embedding vector must be embedded in the obtained new node feature matrix to obtain the final output X' of the growth-aware layer i :

[0083]

[0084] where, is the position embedding matrix The i-th row in the position embedding matrix is obtained by summing the position embedding with the feature matrix.

[0085] Step 8: Through the foregoing steps, the structural growth awareness of the dynamic graph is completed, and then structural dynamic awareness needs to be performed, that is, the multiple node vector matrices obtained are processed using self-attention.

[0086] First, for node v, the corresponding vectors in the Z node feature matrices are taken out and spliced to obtain an input of a dynamic awareness layer where Z is the number of static graphs constituting the dynamic graph, and for nodes appearing less than Z times, the default vector is filled to Z times:

[0087] X v =Padding(Concat([X′ 1,v ,X′ 2,v ,…,X′ Z,v ]) (6)

[0088] Step 9: After obtaining the transition state of node v at multiple positions combined with X v , self-attention mechanism is selected for dynamic awareness. The self-attention mechanism calculates the feature interaction of the multiple states of the node, and performs interaction fusion according to the weight coefficient to perceive the node state change process, and completes the dynamic transition mode capture of a single node.

[0089]

[0090] X′ v =e v (X v W k ) (8)where, and are the query matrix and the key matrix, respectively, used to linearly project the input matrix to different spaces; is a scaling coefficient, which is used to solve the problem of gradient disappearance caused by too large values, is the value matrix, which is used to linearly transform the node matrix.

[0091] Step 10: Through the above steps, the item dynamic transfer pattern in the global dynamic graph structure is captured, but in order to ensure information balance, it is still necessary to establish a single session as a local static graph to extract the item static transfer pattern, and the mapping method is as shown in Figure 3

[0092] For a session s = {v s,1 ,v s,2 ,…,v s,l}, first construct it into an undirected graph g l,s = {V s ,E s}, and then use the graph attention mechanism in the graph structure to capture the transfer pattern between items within a single session. Because the main goal of obtaining local static information is to obtain the representation of a single session, the main basis for measuring the difference in influence between neighbor nodes within a single session is the similarity between the neighbor nodes and the overall session:

[0093]

[0094] wherein represents the representation vector of the session, which is obtained by averaging the sum of the corresponding vectors x s,i of all items v s,i in the session. After obtaining the representation vector of the session, the neighbor nodes of the node v s,i are calculated for similarity, and the correlation coefficient between the two nodes is obtained by using a feedforward neural network after splicing the weight of the edge between the two nodes.

[0095]

[0096]

[0097] wherein and are trainable parameter matrices, ⊙ represents Hadamard product, ‖ is a splicing operation, and LeakyRelu is an activation function.

[0098] Step 11: The neighbor information is obtained by using the obtained correlation coefficient to weight and sum the neighbors of the node. Then the neighbor information and the node information are spliced and sent into a feedforward neural network to complete the information aggregation operation of the node, and the node vector containing the first-order neighbor information

[0099]

[0100] x′ s,i= ReLU(w2[x s,i ‖n s,i ]) (14) where, is a trainable weight parameter matrix, and ReLU is an activation function.

[0101] Step 12: By the above operation, the first-order neighbor information of the node is fused into the node to obtain the first-order correlation between the items.

[0102] In order to extract the long-distance dependence between items in the session, the above operation is continued to repeat to complete the fusion of multi-order neighbor information, denoted as:

[0103]

[0104] where, represents the node v s,i The feature vector after fusing the k-1 order neighbor information, by fusing the k order neighbor information The node feature vector Finally, the local static multi-hop information

[0105] Step 13: After processing the global dynamic graph and the local static graph, the global item dynamic transfer pattern X g And the local item static transfer pattern X l .

[0106] In order to make the final interaction prediction, it is necessary to represent the session based on the above two kinds of information.

[0107]

[0108]

[0109] S = w s [s l ‖s g ] (18) where, s l and s g respectively represent the vector representation obtained by summing and averaging the local and global feature vectors of the nodes in the session, and the vector representation is obtained by concatenating the two vectors and converting them through the parameter matrix w s .

[0110] Step 14: Multiply the vector representation S obtained and the matrix containing all item feature vectors, and normalize it using the softmax function to obtain the score vector of all items

[0111]

[0112] At the same time, the cross-entropy function is selected as the loss function, and the model is trained by cooperating with the time-based back propagation algorithm:

[0113]

[0114] Wherein, y i represents the real label of the item v i . The model reduces the gap between the predicted value and the real value by continuous learning to obtain better recommendation effect.

[0115] According to the operation process of the above steps, the session recommendation can be realized, and the possible next item for each session is predicted.

[0116] After the operation of the above steps, the accuracy (Precision) and the average reciprocal rank (MRR) of the recommendation result obtained by the method (recommendation method based on sequence growth perception model) proposed in the application are shown in the last row of Table 1.

[0117] In order to illustrate the recommendation effect of the application, under the same conditions, the POP, S-POP, Item-KNN, BPR-MF, GRU4Rec, NARM, STAMP, SR-GNN, GCE-GNN and LESSR methods are respectively used for comparison with the same experimental data set.

[0118] Table 1 Comparison of effects of session recommendation methods

[0119]

[0120] The following conclusions can be obtained from Table 1: the session recommendation method based on the sequence growth perception model realizes higher accuracy and average reciprocal rank than the existing session recommendation methods, which verifies the effectiveness of the application.

Claims

1.A method for session recommendation based on a sequence growth perception model, characterized in that, Comprising the following steps: Step 1: Construct a global dynamic graph using the session; Firstly, all sessions are divided according to the established rules, and the growth state obtained by division is established as a discrete dynamic graph structure, called a global dynamic graph; wherein all sessions refer to all session sequences in the data set; The established rule for dividing the session is to divide each session into multiple growth states by using a division window with an initial size of two and growing according to a gradient with a size of one; The discrete dynamic graph is a specific dynamic graph representation form, and the dynamic nature of the graph structure is embodied by multiple sequentially arranged nodes and edges that are not completely the same static graph. Through the evolution of this graph structure, the growth process of the sequence is explicitly modeled; A static graph is composed of the same growth state of several sessions, and each node in it represents an item. The adjacent edge between two nodes represents the existence of adjacent interaction order between the two items, and the number of edges between the two nodes represents the number of adjacent appearances of the two items; Step 2: Construct a local static graph using a single session; Wherein, a single session refers to an independent session sequence and does not contain interactions with other sessions, i.e., a local level of information; the established static graph is a kind of undirected graph structure, in which the nodes correspond to items, and the edges correspond to the adjacent relationship between two items, and the number of edges is consistent with the number of adjacent appearances of the two items; Step 3: Structure growth perception of the global dynamic graph; For a static graph structure, first obtain the node vector matrix through the embedding layer, and then complete the information propagation and aggregation operation between nodes according to the adjacency matrix, thereby updating the node vector matrix; when performing the information propagation and aggregation operation, the attention mechanism is used to calculate the correlation coefficient of the node and its neighbor, and then the correlation coefficient is weighted and summed to complete the targeted information propagation and aggregation; the final vector matrix contains the corresponding graph structure information; Repeat the above structure information extraction operation to obtain a node vector matrix corresponding to the number of static graphs; Embed the corresponding position vector based on the node vector matrix to obtain a node vector matrix corresponding to each static graph containing structure information and order information as the growth perception output; Step 4: Structure dynamic perception of the global dynamic graph; After obtaining the multiple node vector matrices output in step 3, in order to capture the overall structure dynamic change, the self-attention mechanism is used to associate and fuse the multiple node vector matrices to obtain a fused global dynamic transition mode vector matrix; Before fusion, according to the number of nodes in the last node vector matrix, other matrices are padded to the same size, and the vector of the padded node is a zero vector; Step 5: Capture the static transition mode in the local static graph; Use the graph attention mechanism to learn the item transition mode in a single session. The local static information will be fused with the global dynamic information to finally obtain a balanced information session representation for predicting the next interaction; First, the node vector matrix of the local static graph is obtained through the embedding layer, wherein the node vector is the vector corresponding to the item in the local static graph; the session average representation is calculated, which is obtained by averaging the sum of all item node vectors in the local static graph; the similarity between each node vector and the session average representation is calculated to obtain the pair-wise correlation coefficient, and the correlation coefficient is normalized; based on the normalized correlation coefficient, attention weight is assigned to each node to ensure that nodes similar to the session are assigned greater attention weight; information propagation and aggregation between nodes are completed according to the attention weight to obtain the local static transition mode; Step 6: based on the global dynamic transition mode and the local static transition mode, the session representation is obtained; Step 4 obtains the fused vector matrix as the global dynamic transition mode, and step 5 obtains the local static transition mode from a single session; the two are fused through the full connection layer to obtain the final session representation, which considers both global and local level information and dynamic and static transition modes; Step 7: use the session representation to interact with the prediction; Multiply the session representation obtained in step 6 with the vector matrix composed of all items, and then use the softmax function to process it to obtain the probability distribution of all items becoming the next interaction; select the top K items with the largest probability values as the recommendation list, and the value of K is set according to the accuracy requirement; Step 8: use the cross-entropy loss function to train the model; Use the cross-entropy between the real probability distribution and the predicted probability distribution obtained in step 7 as the loss function to train the model, and complete the session recommendation. 2.The method of claim 1, wherein, In step 3, when performing information propagation and aggregation operations, the attention mechanism is used to calculate the correlation coefficient of the node and its neighbors, and then the weighted sum is performed according to the correlation coefficient to complete targeted information propagation and aggregation. 3.The method of claim 1, wherein, In step 4, the vector of the padding node is a zero vector.

Images