Bidirectional Cross-Domain Session Recommendation Method Based on GCNsformer Hybrid Network and Multi-Channel Semantics

Abstract

The present invention relates to a bidirectional cross-domain session recommendation method based on a GCNsformer hybrid network and multi-channel semantics, comprising the following steps: First, obtain session data in domain A and domain B respectively, and establish corresponding global session graphs according to the session data in each domain; Second, according to the global session graphs, use the item embedding representation layer to calculate the node vector representations in each domain respectively; Third, use the multi-channel semantics module layer to calculate complex cross-domain features; Fourth, adopt an enhanced gating mechanism to adaptively combine node vector representations of different types; Step Five, calculate the interaction probabilities for all candidate nodes and generate recommendation lists in different domains simultaneously. The present invention improves the comprehensiveness of session representations in each domain; alleviates the data sparsity problem and the timeliness of dual-domain session recommendation, greatly improves the generalization ability and recommendation recall rate of the bidirectional cross-domain session recommendation model, and realizes the purpose of solving users' commodity demands according to users' cross-domain interaction behaviors.

Description

Technical Field

[0001] The present invention belongs to the technical field of two-way cross-domain session recommendation, and particularly relates to a two-way cross-domain session recommendation method based on a GCNsformer hybrid network and multi-channel semantics. Background Art

[0002] With the continuous development of the Internet and information communication technologies and the popularization of mobile devices, recommendation systems have become the main tool for information services due to their powerful function of accurately linking information and users in massive data. However, in recent years, with the rapid development of Internet enterprises, the data barriers between industries are gradually disappearing, and the session recommendation algorithms based on single-domain interaction information and user preferences are invalid. The birth of the two-way cross-domain serialization recommendation algorithm only needs to predict the items that users are really interested in each domain respectively according to the interaction behavior records of users in different domain sessions, which has played a high application value.

[0003] Graph Convolutional Network (GCN) has been widely used in various types of recommendation algorithms due to its performance in constructing complex interaction patterns. It can be found in many recommendation applications such as social recommendation and spatio-temporal point-of-interest recommendation. Most of the existing relevant research results construct the session sequence structure in the form of a graph structure and show good recommendation performance.

[0004] Scholars from Xidian University proposed a multi-dimensional two-way attention transfer network in 2021. It first uses a method combining a graph structure and a recurrent neural network to learn the item representations in each session, and then uses a mapping modeling method based on the attention mechanism to extract the representation information of users in different sessions. Although the graph structure-based model shows good adaptability in the field of two-way cross-domain session recommendation, the problems of this method are as follows: 1. The existing two-way cross-domain session recommendation methods cannot model the structural relationship in space and the long-distance dependence relationship in time between interactive items at the same time. When focusing on the spatial structure relationship between interactive items, the long-distance dependence relationship inherent in the session sequence itself cannot be well modeled. 2. Due to the relatively complex semantic characteristics in the two-way cross-domain session recommendation scenario and the complex cross-domain feature interaction, the existing methods do not capture the cross-relationship of features completely, resulting in the generated session representation still having poor fitting, and the session recommendation results in different domains have problems of poor accuracy and weak generalization. This causes the session recommendation system to be unable to accurately provide corresponding recommendations for users, reducing the system function and the user's satisfaction with the system at the same time. Summary of the Invention

[0005] The present invention provides a bidirectional cross-domain session recommendation method based on a GCNsformer hybrid network and multi-channel semantics. Aiming at the problem that existing bidirectional cross-domain session recommendation methods cannot simultaneously model the structural relationship in space and the long-distance dependence relationship in time among interactive items, the generated session representation has poor fitting, resulting in poor accuracy and weak generalization in the session recommendation results of different domains.

[0006] To achieve the object of the present invention, a bidirectional cross-domain session recommendation method based on a GCNsformer hybrid network and multi-channel semantics provided by the present invention includes the following steps:

[0007] Step 1: Obtain the session data of domain A and domain B respectively, establish corresponding global session graphs according to the session data in each domain, and obtain the connection relationships of all nodes in the graphs;

[0008] Step 2: According to the global session graph, use the item embedding representation layer to calculate the node vector representations in each domain respectively. Based on the node vector representations, obtain the session local preferences and global preferences of domain A and domain B respectively according to the current user session data. The establishment process of the item embedding representation layer is as follows: First, construct the transfer relationship between items in the session in the form of a graph structure. Second, based on the message passing mechanism, perform node feature aggregation according to the neighbor node feature representations. Subsequently, use the aggregated node representations as the input of the transformer model;

[0009] Step 3: Based on the node vector representations and the session local preferences and global preferences of domain A and domain B, use the multi-channel semantics module layer to calculate complex cross-domain features. The establishment process of the multi-channel semantics module layer is as follows: According to the interaction records of the same user in different domains, use the contrast learning method to learn the common features of the user under cross-domain interaction; Use the mutual attention mechanism to learn the mutual enhancement features between bidirectional cross-domain sessions, and use the adversarial mechanism to obtain the unique features between bidirectional cross-domain sessions;

[0010] Step 4: Based on the node vector representations and the complex cross-domain features, use the enhanced gating mechanism to adaptively combine different types of node vector representations and calculate the session vector representations of different domains;

[0011] Step 5: Calculate the interaction probabilities for all candidate nodes and generate recommendation lists under different domains simultaneously.

[0012] Further, the implementation steps of the above step 3 are as follows:

[0013] (1) Given the interaction sequences of the user in domain A and domain B as and Use the item embedding representation layer to establish the final node vector representations of domain A and domain B respectively and Local and global preferences Furthermore, generate the session representation of the user in different domains and

[0014] (2) Input into the framework of contrastive learning to extract the cross-domain invariant feature expression of the user. The process is expressed as:

[0015]

[0016] where |D| is the number of training samples, N is the number of negative samples, and sim(·) is the similarity calculation function.

[0017] (3) Input and into the mutual attention mechanism to extract the two-way cross-domain mutual enhancement feature expression of the user. The process is expressed as:

[0018]

[0019]

[0020]

[0021]

[0022]

[0023] (4) Input the local and global preferences into the adversarial mechanism to extract the two-way cross-domain unique feature expression of the user. The process is expressed as:

[0024]

[0025] where d(·) and g(·) are the session generator and the domain discriminator respectively.

[0026] Furthermore, the implementation steps of the above step two are as follows:

[0027] (1) Input the session graph G s =(V s , E s ) into the embedding layer to obtain the position vector generated by each node;

[0028] (2) Input the node vector into the multi-head attention mechanism, and calculate the output of the mechanism through three linear transformations and scaled dot-product attention. The process is expressed as:

[0029]

[0030] where H i is the node representation vector constructed for the h-th head attention, and d h is the scaling factor. W i Q 、W i K 、W i V are all trainable parameter matrices.

[0031] (3) After the output of the multi-head self-attention mechanism, a feed-forward neural network is connected. This network generates the final vector representation z i :

[0032] z i = max(0, H i W1 + b1)W2 + b2

[0033] where max(·) is the maximization operation, and W1, W2, b1, and b2 are all trainable parameter matrices.

[0034] (1) For the session s composed of node vectors , the session local preference s l and the global preference s g are respectively represented as:

[0035] α i = softmax(q T · sigmoid(W3z k + W4z i ))

[0036]

[0037]

[0038] Furthermore, in the above step four, the process of the enhanced gating mechanism is represented as:

[0039]

[0040]

[0041]

[0042]

[0043] where g A and g B are respectively the weights corresponding to the session representations in different semantics and different domains, is the vector representation of the session in domain A under the bidirectional cross-domain effect, It is the vector representation of the session in domain B under the two-way cross-domain effect. W A and W B are both trainable parameter matrices.

[0044] Furthermore, the calculation process of the above step five is as follows:

[0045]

[0046]

[0047] where pr A,j is the probability that item v j is the interaction of session s in domain A at the next moment, and pr B,j is the probability that item v j is the interaction of session s in domain B at the next moment;

[0048] ; In the model training process, the cross-loss function is used to calculate the distance between the predicted value and the true value. Then, for session s in domain A, its loss value is expressed as:

[0049]

[0050] Similarly, for session s in domain A, its loss value is expressed as:

[0051]

[0052] Combining the extraction of common features and unique features, the final loss function of the model is:

[0053] Loss = L A + L B + αLoss CL + βLoss GN

[0054] Compared with the prior art, the advantages of the present invention are:

[0055] 1. The present invention mines semantic information of conversation representations from multiple perspectives such as space, time, and two-way collaboration, and simultaneously establishes common features, enhanced features, and unique features across two directions and domains. It not only effectively models the multi-dimensional dynamic preference transfer of the current conversation, but also deeply captures and fuses conversation information in different semantic spaces, significantly improving the comprehensiveness of conversation representations in each domain. Specifically, the present invention constructs a global conversation graph based on the global conversation information in different domains, and uses the GCNsformer hybrid network in the item embedding representation layer to encode items from the perspectives of spatial structure information and time dependence relationships, generating the final vector representations of each node; uses contrastive learning, mutual attention mechanism, and adversarial mechanism in the multi-channel semantic module layer to obtain complex feature cross-information in different domains in the two-way cross-domain conversation recommendation scenario; adopts an enhanced gating mechanism to obtain the final conversation representations in different domains, and then combines the candidate item representations in each domain to calculate the probability that the user may interact with them at the next moment, and generates recommendation information for the next interaction item, thereby achieving the conversation recommendation effect. Among them, different cross-domain common features or unique features are introduced through weights during the training process of the model for recommendation.

[0056] 2. Aiming at the two-way cross-domain conversation recommendation scenario, the present invention effectively models the dynamic preference transfer mode and introduces two-way collaboration information through the designed item embedding representation layer and multi-channel semantic module layer, alleviating the data sparsity problem in conversation recommendation and the timeliness problem of two-domain conversation recommendation, greatly improving the generalization ability and recommendation recall rate of the two-way cross-domain conversation recommendation model, thereby achieving the purpose of solving users' commodity demands according to users' cross-domain interaction behaviors. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 It is a schematic diagram of the connection matrix A of the present invention;

[0058] Figure 2 It is a structural diagram of the GCNsformer hybrid network of the present invention;

[0059] Figure 3 It is a flowchart of the method of the present invention;

[0060] Figure 4 It is a structural diagram of the method of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0061] In order to make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of the present invention, which are only used to illustrate the present invention but not to limit the scope of the present invention.

[0062] The present invention designs a bidirectional cross - domain session recommendation method based on a GCNsformer hybrid network and multi - channel semantics. By spatio - temporal modeling of bidirectional cross - domain session sequences and mining of multi - channel semantic information in bidirectional cross - domains, the final session representation has the ability to construct a multi - dimensional dynamic preference transfer pattern and establish a complete user preference representation, so as to solve the problems of poor recommendation accuracy and weak model generalization still existing in existing session recommendations. Specifically, in implementation, the GCNsformer hybrid network is used to learn the embedded representations of interaction items within each domain; based on the intra - domain local session graph structure and the inter - domain global session graph structure, the multi - channel semantic module proposed in this paper is used to extract the common features, mutually enhanced features, and unique features in the bidirectional cross - domain session recommendation scenario respectively; finally, based on the gating mechanism, the session features in each domain are adaptively integrated with dynamic weights respectively to form the final session representation under different domains.

[0063] See Figure 3 and Figure 4 , a bidirectional cross - domain session recommendation method based on a GCNsformer hybrid network and multi - channel semantics provided by the present invention includes the following steps:

[0064] Step 1: Obtain the session data of domain A and domain B respectively, establish corresponding global session graphs according to the session data in each domain, and obtain the connection relationships of all nodes in the graphs:

[0065] For the bidirectional cross - domain session recommendation scenario, respectively use V d ={v1, v2,..., v m} to represent the only nodes in the single - domain session. Given the user interaction records, first establish a session sequence according to the chronological relationship of the interaction records; then, for the session information in each domain, based on the relevant theories of graph theory and all user interaction records within the domain, use the graph structure to construct the intra - domain global session graph G s =(V s , E s ) to represent the complex transformation relationship between interaction items. See Figure 1 , where V s represents the nodes in the global session graph, and E s represents the edges in the global session graph. Any edge (v i , v i+1 ) means that the user interacted with v i after clicking v i+1 ; finally, according to the global session graph structure, construct a connection matrix A to represent the connection relationship between nodes. Considering that there may be multiple connections between nodes, in order to eliminate the situation of inconsistent dimensions, the edge weights are normalized.

[0066] Step 2: According to the global session graph, use the item embedding representation layer to calculate the node vector representations in each domain respectively. Based on the node vector representations, obtain the session local preferences and global preferences in Domain A and Domain B respectively according to the current user session data. The establishment process of the item embedding representation layer is as follows: First, construct the transfer relationship between items in the session in the form of a graph structure. Second, based on the message passing mechanism, perform node feature aggregation according to the neighbor node feature representations. Subsequently, use the aggregated node representations as the input of the transformer model; see Figure 2 , and the specific implementation steps are as follows:

[0067] (1) Input the session graph G s =(V s , E s ) into the embedding layer, and use the random embedding method to represent the node features as low-dimensional dense vectors e i = Embedding_layer(num_v, emb_dim), where num_v is the number of nodes in the global session graph, and emb_dim is the dimension of the node feature vector;

[0068] (2) The l-layer information propagation process between neighbor nodes and the current node under the action of the connection matrix A is expressed as: where N(v i ) is the set of v i nodes and their l-order neighbor nodes, and the node vector representations corresponding to each node in any session s are

[0069] (3) For the node vector in session s, in order to characterize the sequential relationship between items, use the linear transformation of the sin and cos functions to generate a position vector for each node, and the process is expressed as:

[0070]

[0071]

[0072] where p is the position of the i-th item in the session, d is the dimension of the item vector, and respectively represent the position embedding vectors of the nodes.

[0073] (4) Input the node vector into the multi-head attention mechanism, and calculate the output of this mechanism through three linear transformations and scaled dot-product attention. The process is expressed as:

[0074]

[0075] where H iThe node representation vector constructed for the h-th head attention, d h is the scaling factor. W i Q 、W i K 、W i V are all trainable parameter matrices.

[0076] (5) To increase the non-linearity of the GCNsformer hybrid network, a feed-forward neural network is connected after the output of the multi-head self-attention mechanism. This network generates the final vector representation z i :

[0077] z i =max(0, H i W1 + b1)W2 + b2

[0078] where max(·) is the maximization operation, and W1, W2, b1, and b2 are all trainable parameter matrices.

[0079] (6) For the session s composed of node vectors , the session local preference s l and the global preference s g are respectively represented as:

[0080] α i =softmax(q T ·sigmoid(W3z k +W4z i ))

[0081]

[0082]

[0083] Step 3: Based on the node vector representation and the session local preferences and global preferences in domains A and B, use the multi-channel semantic module layer to calculate complex cross-domain features. The key is to establish the multi-channel semantic module layer. The establishment process of the multi-channel semantic module layer is as follows: According to the interaction records of the same user in different domains, use the contrastive learning method to learn the common features of the user under cross-domain interaction; adopt the mutual attention mechanism to learn the mutual enhancement features between two-way cross-domain sessions, and adopt the adversarial mechanism to obtain the unique features between two-way cross-domain sessions. The specific implementation steps are as follows:

[0084] (1) Given the interaction sequences of the user in domains A and B as and Use the item embedding representation layer to respectively establish the final vector representations of the nodes in domains A and B and Local and global preferences Furthermore, generate the conversation representation of the user in different domains and

[0085] (2) To obtain common features, input into the framework of contrastive learning. By reducing the differences in the conversation representations of the user in different domains, the purpose of extracting the cross-domain invariant feature expression of the user is achieved. The process is expressed as:

[0086]

[0087] where |D| is the number of training samples, N is the number of negative samples, and sim(·) is the similarity calculation function.

[0088] (3) To obtain mutual enhancement features, input and into the mutual attention mechanism. By establishing the correlation between the interactive items in different domains, the purpose of extracting the two-way cross-domain mutual enhancement feature expression of the user is achieved. The process is expressed as:

[0089]

[0090]

[0091]

[0092]

[0093]

[0094] (4) To obtain unique features, input the local and global preferences into the adversarial mechanism. By distinguishing the vector representation of the conversation in the current domain and the cross-domain global vector representation, the purpose of extracting the two-way cross-domain unique feature expression of the user is achieved. The process is expressed as:

[0095]

[0096] where d(·) and g(·) are the conversation generator and the domain discriminator respectively.

[0097] Step 4: Based on the node vector representation and the complex cross-domain features, adopt an enhanced gating mechanism to adaptively combine the node vector representations of different types, and calculate the conversation vector representations of different domains. The process of the enhanced gating mechanism is expressed as:

[0098]

[0099]

[0100]

[0101] '

[0102] where g A and g B are the weights corresponding to the session representations under different semantics and different domains respectively, is the vector representation of the session in domain A under the bidirectional cross - domain effect, is the vector representation of the session in domain B under the bidirectional cross - domain effect, W A and W B are both trainable parameter matrices.

[0103] Step 5: Calculate the interaction probabilities for all candidate nodes and generate recommendation lists under different domains simultaneously: In each domain, the prediction of the possible interaction items for the next moment of each session is achieved by calculating the probability distribution of the session representation and all candidate items in each domain. The specific calculation process is as follows:

[0104]

[0105]

[0106] where pr A,j is the probability that item v j interacts with session s in domain A at the next moment, and pr B,j is the probability that item v j interacts with session s in domain B at the next moment. Then, the probability values are sorted from largest to smallest respectively, and the item with a larger probability is more likely to be recommended.

[0107] Since the session recommendation problem essentially belongs to a classification task, the cross - entropy loss function is used in the model training process to calculate the distance between the predicted value and the true value. Then, for session s in domain A, its loss value is expressed as:

[0108]

[0109] Similarly, for session s in domain A, its loss value is expressed as:

[0110]

[0111] Combining the extraction of common features and unique features, the final loss function of the model is:

[0112] Loss = L A + L B + αLossCL +β Loss GN 。

[0113] The above is a description of the specific implementation of the present invention, rather than a limitation thereof. Those skilled in the relevant technical field can also make various equivalent technical solutions without departing from the scope of the present invention. Therefore, all equivalent technical solutions should be included in the scope of patent protection of the present invention.

Claims

1. A bidirectional cross - domain session recommendation method based on the GCNsformer hybrid network and multi - channel semantics, characterized in that: It includes the following steps: Step 1: Obtain the session data in fields A and B respectively, establish corresponding global session graphs according to the session data in each field, and obtain the connection relationships of all nodes in the graphs; Step 2: According to the global session graphs, use the item embedding representation layer to calculate the node vector representations in each field respectively; Based on the node vector representations, obtain the session local preferences and global preferences in fields A and B respectively according to the current user session data. The establishment process of the item embedding representation layer is as follows: First, construct the transfer relationship between items in the session in the form of a graph structure. Second, based on the message passing mechanism, perform node feature aggregation according to the neighbor node feature representations. Subsequently, use the aggregated node representations as the input of the transformer model; Step 3: Based on the node vector representations and the session local preferences and global preferences in fields A and B, use the multi-channel semantic module layer to calculate complex cross-domain features. The establishment process of the multi-channel semantic module layer is as follows: According to the interaction records of the same user in different domains, use the contrastive learning method to learn the common features of the user under cross-domain interaction; Adopt the mutual attention mechanism to learn the mutual enhancement features between two-way cross-domain sessions, and adopt the adversarial mechanism to obtain the unique features between two-way cross-domain sessions; Step 4: Based on the node vector representations and the complex cross-domain features, adopt the enhanced gating mechanism to adaptively combine different types of node vector representations and calculate the session vector representations in different fields; Step 5: Calculate the interaction probabilities for all candidate nodes and generate recommendation lists in different domains simultaneously.

2. A bidirectional cross - domain session recommendation method based on the GCNsformer hybrid network and multi - channel semantics according to claim 1, characterized in that: The implementation steps of Step 3 are as follows: (1) Given that the interaction sequences of a user in domain A and domain B are and Use the item embedding representation layer to respectively establish the final vector representations of the nodes in domain A and domain B and Local and global preferences Furthermore, generate the session representations of the user in different domains and (2) Input into the contrastive learning framework to extract the cross-domain invariant feature representation of the user, and the process is expressed as: where |D| is the number of training samples, N is the number of negative samples, and sim(·) is the similarity calculation function; (3) Input and into the mutual attention mechanism to extract the user's bidirectional cross-domain mutual enhancement feature representation. The process is expressed as: (4) Input local and global preferences into the adversarial mechanism to extract the user's two-way cross-domain unique feature expression. The process is expressed as: where d(·) and g(·) are the session generator and the domain discriminator respectively.

3. A bidirectional cross-domain session recommendation method based on a GCNsformer hybrid network and multi-channel semantics according to claim 2, characterized in that: The implementation steps of Step 2 are as follows: (1) Input the session graph G s = (V s , E s ) into the input embedding layer to obtain the position vectors generated by each node; (2) Input the node vector into the multi-head attention mechanism. Calculate the output of this mechanism through three linear transformations and scaled dot-product attention. The process is expressed as: Among which H i is the node representation vector constructed for the h-th head attention, and d h is the scaling factor; W i Q , W i K , W i V are all trainable parameter matrices; (3) After the output of the multi-head self-attention mechanism, a feed-forward neural network is connected. This network generates the final vector representation z of item node i through linear and activation functions i : z i = max(0, H i W1 + b1)W2 + b2 where max(·) is the maximization operation, and W1, W2, b1, and b2 are all trainable parameter matrices; (4) For a session s composed of a knot vector the session local preference s l and the global preference s g are respectively represented as: α i = soft max(q T · sigmoid(W3z k + W4z i )) 4. A bidirectional cross - domain session recommendation method based on a GCNsformer hybrid network and multi - channel semantics according to claim 3, characterized in that: In Step 4, the process of the enhanced gating mechanism is expressed as: where g A and g B are the weights corresponding to the session representations under different semantics and different domains respectively, is the vector representation of the session in domain A under the bidirectional cross-domain effect, is the vector representation of the session in domain B under the bidirectional cross-domain effect, W A and W B are both trainable parameter matrices.

5. A bidirectional cross-domain session recommendation method based on a GCNsformer hybrid network and multi-channel semantics according to any one of claims 1-4, characterized in that: The calculation process of Step 5 is as follows: where pr A,j is the probability that the item v j interacts as the interaction of the session s in the next moment in field A, and pr B,j is the probability that the item v j interacts as the interaction of the session s in the next moment in field B; In the model training process, the cross-loss function is used to calculate the distance between the predicted value and the true value. For the session s in field A, its loss value is expressed as: Similarly, for the session s in field A, its loss value is expressed as: Combined with the extraction of common features and unique features, the final loss function of the model is: Loss=L A +L B +αLoss CL +βLoss GN 。

Images