Sequence recommendation method based on multi-behavior dynamic graph

Abstract

The invention belongs to the technical field of sequence recommendation, and particularly relates to a sequence recommendation method based on a multi-behavior dynamic graph. Comprising the following steps: constructing a multi-behavior dynamic graph according to a multi-behavior interaction sequence of a user and dividing the multi-behavior dynamic graph into a plurality of specific behavior sub-graphs; processing the specific behavior sub-graph to obtain a node embedding representation and a relation embedding representation; processing the relation embedding representation to obtain a general correlation embedding representation; time sequence information under the specific behavior is fused into nodes of the specific behavior subgraph, and a fused user node embedding expression and a fused project node embedding expression are obtained; fusing the general correlation embedding representation and the user node embedding representation to obtain a user personalized behavior correlation embedding representation; updating the fusion project node embedding representation and aggregating the updating result and the fusion user node embedding representation to obtain two final node embedding representations; processing the final two node embedding representations by adopting a feedforward neural network to obtain a project recommendation result; according to the invention, the recommendation accuracy is improved.

Description

Technical Field

[0001] The present invention belongs to the technical field of sequence recommendation, and in particular relates to a sequence recommendation method based on multi-behavior dynamic graphs. Background Art

[0002] Traditional recommendation systems often use static methods to analyze the interaction between users and products. This model mainly captures the general preferences of users. However, in real-world applications, user behavior changes continuously. Their interaction methods, preferences, and the popularity of products will change over time. In order to cope with this dynamic nature, sequential recommendation systems were introduced.

[0003] The sequential recommendation system regards the interaction between users and products as a continuous sequence, and uses the correlation within the sequence to deeply understand the user's current and recent interests. By analyzing the continuity of user behavior, this system can gain a deeper insight into the immediate state and recent trends of user interests. Therefore, the sequential recommendation system breaks through the limitations of traditional recommendation models and provides a more accurate way to reflect the dynamic changes of user interests over time.

[0004] Given the significant practical value of sequential recommendation systems, research on sequential recommendation issues is booming, and various innovative solutions have emerged. In this field, some methods based on the idea of ​​Markov chains have been introduced. However, under the assumption that each behavior is independent, the independent combination of components in the past has limited the accuracy of predictions. To overcome this limitation, deep learning-based methods have emerged, among which the main deep neural network models used include graph neural networks (GNNs), recurrent neural networks (RNNs), and convolutional neural networks (CNNs). These deep learning methods can better characterize the dynamic interaction between users and items and learn effective user preference representations. Numerous experiments have confirmed the advantages of these methods, but the following problems still exist:

[0005] Current sequence recommendation methods generally focus on the user's target behavior data, thus forming a user-item interaction sequence. However, in the real world, users usually interact with items in a multi-behavior serialized manner. For example, on e-commerce platforms, interaction behaviors include clicking, adding to favorites, adding to shopping carts, and purchasing. The information in multi-behavior interaction data often provides users with more fine-grained interest preferences. In addition, current multi-behavior sequence recommendation methods do not fully consider the temporal information in the interaction sequence and ignore the dynamic changes of users' interest preferences. Summary of the invention

[0006] In view of the shortcomings of the prior art, the present invention proposes a sequence recommendation method based on a multi-behavior dynamic graph, the method comprising: obtaining a multi-behavior interaction sequence of a user to be recommended; inputting the multi-behavior interaction sequence into a trained item recommendation model to obtain an item recommendation result;

[0007] The training process of the item recommendation model includes:

[0008] S1: Obtain the user's historical behavior data and extract multi-behavior interaction sequences from the historical behavior data;

[0009] S2: Construct a multi-behavior dynamic graph based on the multi-behavior interaction sequence, and divide the multi-behavior dynamic graph into multiple specific behavior subgraphs;

[0010] S3: Initialize the node embedding representation of the specific behavior subgraph to obtain the initial node embedding representation; project all the initial node embedding representations into the same feature space to obtain the intermediate node embedding representation;

[0011] S4: Use a multi-behavior encoder to process the intermediate node embedding representation of a specific behavior subgraph to obtain the relational embedding representation of the specific behavior subgraph;

[0012] S5: Use the universal behavior correlation encoder to process the relationship embedding representation of any two specific behavior subgraphs to obtain the universal correlation embedding representation of every two specific behavior subgraphs;

[0013] S6: Fuse the temporal information under the specific behavior into the nodes of the specific behavior subgraph to obtain the fused user node embedding representation and the fused item node embedding representation under the specific behavior subgraph;

[0014] S7: Fusion of the universal correlation embedding representation of each two specific behavior subgraphs and fusion of the user node embedding representation to obtain the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user;

[0015] S8: Based on the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user, the fusion project node embedding representation is processed using the attention mechanism to obtain an updated fusion project node embedding representation;

[0016] S9: Aggregate the updated fused project node embedding representation under each specific behavior subgraph to obtain the final project node embedding representation; aggregate the fused user node embedding representation under each specific behavior subgraph to obtain the final user node embedding representation;

[0017] S10: input the final project node embedding representation and the final user node embedding representation into a feedforward neural network for processing to obtain a matching score;

[0018] S11: Calculate the total model loss based on the matching score, and adjust the model parameters based on the total model loss. When the loss function converges, the model training is completed.

[0019] Furthermore, the multi-behavior dynamic graph is represented as:

[0020] G={(u,v,b,t,o v,u ,o u,v )|u∈U,v∈V,b∈B}

[0021] Among them, G represents a multi-behavior dynamic graph, U represents a user set, V represents a project set, B represents a behavior set, b represents the behavior type of interaction, t represents the timestamp of the interaction, and o represents the time stamp of the interaction. v,u represents the position of item v among all items that user u has interacted with, o u,v represents the position of user u among all users who have interacted with item v;

[0022] Divide the multi-behavior dynamic graph into multiple specific behavior subgraphs as follows:

[0023]

[0024] Among them, G b represents a subgraph that contains only behavior b, U b represents the set of users who have interacted with the project through behavior b, V b represents the set of items that have interacted with the user's behavior b, t represents the timestamp of the interaction, represents the position of item v among all items that user u has interacted with in behavior b, represents the position of user u among all users who have interacted with item v in behavior b.

[0025] Furthermore, the formula for obtaining the relational embedding representation of a specific behavior subgraph is:

[0026] M u =Mean(dropout(H u ))

[0027] M v =Mean(dropout(H v ))

[0028] h b =σ(W2[M u ‖M v ]+b0)

[0029] Among them, H u is a specific behavior subgraph G b The intermediate user nodes in H represent the set, vis a specific behavior subgraph G b The middle item node in h represents a collection; b represents the relational embedding representation of a specific behavior subgraph; σ represents the activation function; ∥ represents the vector concatenation operation; M u represents the aggregated user representation, M v represents the aggregated item representation; b0 represents the first bias; W2 represents the third parameter matrix.

[0030] Furthermore, the formula for obtaining the universal correlation embedding representation of every two specific behavior subgraphs is:

[0031]

[0032] in, represents the universal correlation embedding representation of the i-th specific behavior subgraph and the j-th specific behavior subgraph, represents the relation embedding representation of the i-th specific behavior subgraph, represents the relational embedding representation of the j-th specific behavior subgraph, b1 represents the second bias, σ represents the activation function; ∥ represents the vector concatenation operation, and W3 represents the fourth parameter matrix.

[0033] Furthermore, the process of obtaining the fused user node embedding representation and the fused project node embedding representation under the specific behavior subgraph includes:

[0034] The relation embedding representation of the specific behavior subgraph, the user's intermediate node embedding representation and the project's intermediate node embedding representation are integrated to obtain a first attention score; the first attention score is normalized to obtain a first attention weight; the fused user node embedding representation is calculated based on the first attention weight and the project's intermediate node embedding representation;

[0035] The relation embedding representation of the specific behavior subgraph, the intermediate node embedding representation of the project and the intermediate node embedding representation of the user are integrated to obtain the second attention score; the second attention score is normalized to obtain the second attention weight; the fused project node embedding representation is calculated based on the second attention weight and the intermediate node embedding representation of the user.

[0036] Furthermore, the formula for obtaining the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user is:

[0037]

[0038] in, represents the personalized behavior correlation rate between the i-th specific behavior subgraph and the j-th specific behavior subgraph of user u; A universal correlation embedding representation representing the i-th specific behavior subgraph and the j-th specific behavior subgraph; represents the personalized behavior correlation embedding representation of the i-th specific behavior subgraph and the j-th specific behavior subgraph of user u; σ represents the activation function, and W6 and W7 are distributed as the seventh and eighth parameter matrices.

[0039] Furthermore, the formula for updating the embedding representation of the fusion project node is obtained as follows:

[0040]

[0041] Among them, q i represents the query vector of the i-th item that user u interacts with, Indicates b (u,i) The fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph, b represents user u (u,i) Behavior subgraph and b (u,j) Personalized behavior relevance embedding representation of behavior subgraph, k j represents the key vector of the jth item that user u interacted with, Indicates b (u,j) The fused item node embedding representation of the jth item interacted with user u under the behavior subgraph, v j represents the value vector of the jth item interacting with user u, A[i,j] represents the attention weight of the jth item interacting with user u to the ith item interacting with user u, softmax represents the softmax activation function, N u Represents the set of items that the user has interacted with, Indicates b (u,i) The updated fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph.

[0042] Furthermore, the formula for obtaining the project matching score is:

[0043]

[0044] Among them, score uv Indicates the matching score between user u and item v, score u represents the matching score vector of user u, represents the final user node embedding representation, represents the final project node embedding representation, W pre represents the prediction parameter matrix, softmax represents the softmax activation function, Represents the normalized matching score vector.

[0045] Furthermore, the formula for calculating the total loss of the model is:

[0046]

[0047] Among them, LOSS represents the total loss of the model, S represents the sequence set; y uv Indicates whether the next interaction of user u is item v, is the normalized matching score between user u and item v, θ represents all model parameters, ‖·‖2 represents the L2 norm; λ is the regularization strength control coefficient.

[0048] The beneficial effects of the present invention are as follows: the present invention simultaneously takes into account the multi-behavior information and timing information of the user-project multi-behavior interaction sequence, uses a multi-behavior dynamic graph to store the interaction sequence data, uses methods such as the attention mechanism to learn and update user embedding, project embedding, and behavior embedding, and then sends the final user representation and the candidate items to be predicted into a feedforward neural network to calculate the matching score. According to the matching score, more accurate recommendation objects can be provided to the user, thereby improving the accuracy of the recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

[0049] Figure 1 This is a training flow chart of the project recommendation model in the present invention;

[0050] Figure 2 This is a diagram of the training framework of the project recommendation model in the present invention. DETAILED DESCRIPTION

[0051] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0052] The present invention proposes a sequence recommendation method based on a multi-behavior dynamic graph, the method comprising:

[0053] Obtain the multi-behavior interaction sequence of the user to be recommended; input the multi-behavior interaction sequence into the trained item recommendation model to obtain the item recommendation result.

[0054] like Figure 1 , Figure 2 As shown in the figure, the training process of the item recommendation model includes:

[0055] S1: Obtain the user's historical behavior data and extract multi-behavior interaction sequences from the historical behavior data.

[0056] Get the user's historical behavior data, which includes different users' past behaviors on different items. For example, in product recommendations, each user's interactive behavior data on different products, such as clicks, favorites, and purchases; extract multi-behavior interaction sequences from the historical behavior data. The multi-behavior interaction sequence {u1, v1 click, t1} represents the interactive behavior of user u1 clicking on product v1 at time t1.

[0057] The user's next interaction item information is used as a label. For example, if the next interaction item of user u is item v, then the label y uv is 1 if the value is true, otherwise it is 0.

[0058] S2: Construct a multi-behavior dynamic graph based on the multi-behavior interaction sequence, and divide the multi-behavior dynamic graph into multiple specific behavior sub-graphs.

[0059] According to the multi-behavior interaction sequence, a multi-behavior dynamic graph is obtained, which is expressed as:

[0060] G={(u,v,b,t,o v,u ,o u,v )|u∈U,v∈V,b∈B}

[0061] Among them, G represents a multi-behavior dynamic graph, U represents a user set, V represents a project set, B represents a behavior set, b represents the behavior type of interaction, t represents the timestamp of the interaction, and o represents the time stamp of the interaction. v,u represents the position of item v among all items that user u has interacted with, o u,v represents the position of user u among all users who have interacted with item v.

[0062] The multi-behavior dynamic graph is divided into multiple specific behavior subgraphs. The edges and connected nodes of the multi-behavior dynamic graph G that only contain specific behaviors b are taken. The division process is expressed as:

[0063]

[0064] Among them, G b represents a subgraph that contains only behavior b, U b represents the set of users who have interacted with the project through behavior b, V b represents the set of items that have interacted with the user's behavior b, t represents the timestamp of the interaction, represents the position of item v among all items that user u has interacted with in behavior b, represents the position of user u among all users who have interacted with item v in behavior b.

[0065] S3: Initialize the node embedding representation of the specific behavior subgraph to obtain the initial node embedding representation; project all initial node embedding representations into the same feature space to obtain the intermediate node embedding representation.

[0066] Initialize the node embedding representation of a specific behavior subgraph, and use the random walk method for the nodes in the behavior subgraph to obtain the initialized embedding representation of the nodes.

[0067] Considering the heterogeneity of nodes, different types of nodes may exist in different feature spaces, so a type-specific transformation matrix is ​​used to project different types of nodes into the same feature space, expressed as:

[0068] h u =W0x u

[0069] h v =W1x v

[0070] in, h u is the user's intermediate node embedding representation obtained after projection, h v is the embedding representation of the intermediate nodes of the project after projection; x u represents the embedding representation of the user node, x v represents the embedded representation of the project node; d i The dimension after the node represents the projection, d u is the initial user node representation dimension, d v The initialized project node represents the dimension; W0 and W1 represent the first and second parameter matrices respectively.

[0071] S4: Use a multi-behavior encoder to process the intermediate node embedding representation of a specific behavior subgraph to obtain the relational embedding representation of the specific behavior subgraph.

[0072] The multi-behavior encoder generates relation representations from nodes on the relation-specific subgraph, and finally obtains the relation embedding representation of the specific behavior subgraph; first, the behavior subgraph G is aggregated separately b The user representation and item representation in are expressed as:

[0073] M u =Mean(dropout(H u ))

[0074] M v =Mean(dropout(H v ))

[0075] Among them, H u is a specific behavior subgraph G bThe intermediate user nodes in H represent the set, v is a specific behavior subgraph G b The middle item node in M ​​represents a collection; u represents the aggregated user representation, M v Represents the aggregated item representation; Dropout represents the dropout operation, and Mean represents the average operation.

[0076] Then the relation embedding representation of the specific behavior subgraph is generated, which is expressed as:

[0077] h b =σ(W2[M u ‖M v ]+b0)

[0078] Among them, σ represents the activation function; ∥ represents the vector concatenation operation; b0 represents the first bias; W2 represents the third parameter matrix.

[0079] S5: Use the universal behavior correlation encoder to process the relation embedding representation of any two specific behavior subgraphs to obtain the universal correlation embedding representation of every two specific behavior subgraphs.

[0080] A universal behavior relevance encoder is used to learn any set of relation pairs, and a universal relevance embedding representation of any two relations is obtained:

[0081]

[0082] in, A universal correlation embedding representation representing the i-th specific behavior subgraph and the j-th specific behavior subgraph; is a set of behavior relationship pairs, represents the relation embedding representation of the i-th specific behavior subgraph, represents the relational embedding representation of the j-th specific behavior subgraph, b1 represents the second bias, σ represents the activation function; ∥ represents the vector concatenation operation, and W3 represents the fourth parameter matrix.

[0083] S6: Fusion the temporal information under the specific behavior into the nodes of the specific behavior subgraph to obtain the fused user node embedding representation and the fused project node embedding representation under the specific behavior subgraph.

[0084] The user's interaction with the project at different times actually reflects the dynamic changes of the user's interest preferences. Therefore, the attention mechanism is used to fuse the temporal information under specific behaviors into the user node, and a new user node representation is obtained, that is, the fused user node embedding representation under the specific behavior subgraph:

[0085]

[0086] in, represents the fused user node embedding representation of user u under a specific behavior subgraph, is the set of all items that user u has interacted with under behavior b; is with Directly related to the position encoding vector for calculating the attention score; is the first attention score of item v for user u under behavior b; yes Attention weights obtained after normalization; is with Directly related to the position encoding vector for information propagation; W4 is the fifth parameter matrix.

[0087] Similarly, the interaction between the project and the user at different times actually reflects the dynamic changes in the project's attributes. Therefore, the attention mechanism is used to fuse the temporal information under specific behaviors into the project node to obtain a new project node representation, namely the fused project node embedding representation:

[0088]

[0089] in, represents the fused item node embedding representation of item v under the feature behavior subgraph, is the set of all users that item v has interacted with under behavior b; is with Directly related to the position encoding vector for calculating the attention score; is the second attention score of user u for item v under behavior b; yes The normalized attention weights; is with Directly related to the position encoding vector for information propagation; W5 is the sixth parameter matrix.

[0090] S7: Fusion of the universal correlation embedding representation of each two specific behavior subgraphs and fusion of the user node embedding representation to obtain the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user.

[0091] Fusion of the universal correlation embedding representation of each two specific behavior subgraphs and the fused user node embedding representation:

[0092]

[0093] in, Represents the personalized behavior correlation rate between the i-th specific behavior subgraph and the j-th specific behavior subgraph of the user; Represents the personalized behavior correlation embedding representation of the user's i-th specific behavior subgraph and the j-th specific behavior subgraph; W6 and W7 are distributed as the seventh and eighth parameter matrices.

[0094] S8: Based on the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user, the attention mechanism is used to process the fusion project node embedding representation to obtain the updated fusion project node embedding representation.

[0095] Use the attention mechanism to fuse personalized behavioral correlations and temporal information into item representations to update item representations:

[0096]

[0097] Among them, q i represents the query vector of the i-th item that user u interacts with, Indicates b (u,i) The fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph, Indicates the user's (u,i) Behavior subgraph and b (u,j) Personalized behavior relevance embedding representation of behavior subgraph, k j represents the key vector of the jth item that user u interacted with, Indicates b (u,j) The fused item node embedding representation of the jth item interacted with user u under the behavior subgraph, v j represents the value vector of the jth item interacting with user u, A[i,j] represents the attention weight of the jth item interacting with user u to the ith item interacting with user u, softmax represents the softmax activation function, N u Represents the set of items that the user has interacted with, Indicates b (u,i) The updated fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph.

[0098] S9: Aggregate the updated fused project node embedding representation under each specific behavior subgraph to obtain the final project node embedding representation; aggregate the fused user node embedding representation under each specific behavior subgraph to obtain the final user node embedding representation.

[0099] The final project node embedding representation is:

[0100]

[0101] The final user node embedding representation is:

[0102]

[0103] S10: Input the final project node embedding representation and the final user node embedding representation into the feedforward neural network for processing to obtain a matching score.

[0104] The final project node embedding representation and the final user node embedding representation are input into the feedforward neural network for processing:

[0105]

[0106] Among them, score uv represents the matching score between user u and item v, W pre Represents the prediction parameter matrix, score u represents the matching score vector of user u, softmax represents the softmax activation function, Represents the normalized matching score vector.

[0107] S11: Calculate the total model loss based on the matching score, and adjust the model parameters based on the total model loss. When the loss function converges, the model training is completed.

[0108] The present invention uses binary cross entropy as the loss function of the recommendation task, i.e., the total model loss, and calculates the total model loss:

[0109]

[0110] Among them, LOSS represents the total loss of the model, S represents the sequence set; y uv is the true label, indicating whether the user's next interaction is item v. If it is v, it takes 1, otherwise it takes 0; is the normalized matching score between user u and item v, θ represents all model parameters, ‖·‖2 represents the L2 norm; λ is the regularization strength control coefficient, which is used to control the regularization strength.

[0111] Adjust the model parameters according to the total loss of the model. When the loss function converges, keep the model parameters, complete the model training, and obtain the trained item recommendation model. Obtain the multi-behavior interaction sequence of the user to be recommended; input the multi-behavior interaction sequence into the trained item recommendation model to obtain the best item recommendation result.

[0112] The above embodiments further illustrate the purpose, technical solutions and advantages of the present invention in detail. It should be understood that the above embodiments are only preferred implementation modes of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made to the present invention within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A sequence recommendation method based on multi-behavior dynamic graph, characterized in that: include: Obtain multiple behavior interaction sequences of the user to be recommended; Input the multi-behavior interaction sequence into the trained item recommendation model to obtain the item recommendation results; The training process of the item recommendation model includes: S1: Obtain the user's historical behavior data and extract multi-behavior interaction sequences from the historical behavior data; S2: Construct a multi-behavior dynamic graph based on the multi-behavior interaction sequence, and divide the multi-behavior dynamic graph into multiple specific behavior subgraphs; S3: Initialize the node embedding representation of the specific behavior subgraph to obtain the initial node embedding representation; project all the initial node embedding representations into the same feature space to obtain the intermediate node embedding representation; S4: Use a multi-behavior encoder to process the intermediate node embedding representation of a specific behavior subgraph to obtain the relational embedding representation of the specific behavior subgraph; S5: Use the universal behavior correlation encoder to process the relationship embedding representation of any two specific behavior subgraphs to obtain the universal correlation embedding representation of every two specific behavior subgraphs; S6: Fuse the temporal information under the specific behavior into the nodes of the specific behavior subgraph to obtain the fused user node embedding representation and the fused item node embedding representation under the specific behavior subgraph; S7: Fusion of the universal correlation embedding representation of each two specific behavior subgraphs and fusion of the user node embedding representation to obtain the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user; S8: Based on the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user, the fusion project node embedding representation is processed using the attention mechanism to obtain an updated fusion project node embedding representation; S9: Aggregate the updated fused project node embedding representation under each specific behavior subgraph to obtain the final project node embedding representation; aggregate the fused user node embedding representation under each specific behavior subgraph to obtain the final user node embedding representation; S10: input the final project node embedding representation and the final user node embedding representation into a feedforward neural network for processing to obtain a matching score; S11: Calculate the total model loss based on the matching score, and adjust the model parameters based on the total model loss. When the loss function converges, the model training is completed.

2. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The multi-behavior dynamic graph is represented as: G={(u,v,b,t,o v，u ,o u，v )|u∈U,v∈V,b∈B} Among them, G represents a multi-behavior dynamic graph, U represents a user set, V represents a project set, B represents a behavior set, b represents the behavior type of interaction, t represents the timestamp of the interaction, and o represents the time stamp of the interaction. v，u represents the position of item v among all items that user u has interacted with, o u，v represents the position of user u among all users who have interacted with item v; Divide the multi-behavior dynamic graph into multiple specific behavior subgraphs as follows: Among them, G b represents a subgraph that contains only behavior b, U b represents the set of users who have interacted with the project through behavior b, V b represents the set of items that have interacted with the user's behavior b, t represents the timestamp of the interaction, represents the position of item v among all items that user u has interacted with in behavior b, represents the position of user u among all users who have interacted with item v in behavior b.

3. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The formula for obtaining the relational embedding representation of a specific behavior subgraph is: M u =Mean(dropout(H u )) M v =Mean(dropout(H v )) h b =σ(W2[M u ||M v ]+b0) Among them, H u is a specific behavior subgraph G b The intermediate user nodes in H represent the set, v is a specific behavior subgraph G b The middle item node in h represents a collection; b represents the relational embedding representation of a specific behavior subgraph; σ represents the activation function; || represents the vector concatenation operation; M u represents the aggregated user representation, M v represents the aggregated item representation; b0 represents the first bias; W2 represents the third parameter matrix.

4. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The formula for obtaining the universal correlation embedding representation of every two specific behavior subgraphs is: in, represents the universal correlation embedding representation of the i-th specific behavior subgraph and the j-th specific behavior subgraph, represents the relation embedding representation of the i-th specific behavior subgraph, represents the relational embedding representation of the j-th specific behavior subgraph, b1 represents the second bias, σ represents the activation function; ∥ represents the vector concatenation operation, and W3 represents the fourth parameter matrix.

5. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The process of obtaining the fused user node embedding representation and the fused item node embedding representation under a specific behavior subgraph includes: The relation embedding representation of the specific behavior subgraph, the user's intermediate node embedding representation and the project's intermediate node embedding representation are integrated to obtain a first attention score; the first attention score is normalized to obtain a first attention weight; the fused user node embedding representation is calculated based on the first attention weight and the project's intermediate node embedding representation; The relation embedding representation of the specific behavior subgraph, the intermediate node embedding representation of the project and the intermediate node embedding representation of the user are integrated to obtain the second attention score; the second attention score is normalized to obtain the second attention weight; the fused project node embedding representation is calculated based on the second attention weight and the intermediate node embedding representation of the user.

6. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The formula for obtaining the personalized behavior correlation embedding representation of each two specific behavior subgraphs of the user is: in, represents the personalized behavior correlation rate between the i-th specific behavior subgraph and the j-th specific behavior subgraph of user u; A universal correlation embedding representation representing the i-th specific behavior subgraph and the j-th specific behavior subgraph; represents the personalized behavior correlation embedding representation of the i-th specific behavior subgraph and the j-th specific behavior subgraph of user u; σ represents the activation function, and W6 and W7 are distributed as the seventh and eighth parameter matrices.

7. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The formula for updating the embedded representation of the fusion project node is: Among them, q i represents the query vector of the i-th item that user u interacts with, Indicates b (u,i) The fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph, b represents user u (u,i) Behavior subgraph and b (u,j) Personalized behavior relevance embedding representation of behavior subgraph, k j represents the key vector of the jth item that user u interacted with, Indicates b (u,j) The fused item node embedding representation of the jth item interacted with user u under the behavior subgraph, v j represents the value vector of the jth item interacting with user u, A[i,j] represents the attention weight of the jth item interacting with user u to the ith item interacting with user u, softmax represents the softmax activation function, N u Represents the set of items that the user has interacted with, Indicates b (u,i) The updated fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph.

8. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The formula for getting the project matching score is: Among them, score uv Indicates the matching score between user u and item v, score u represents the matching score vector of user u, represents the final user node embedding representation, represents the final project node embedding representation, W pre represents the prediction parameter matrix, softmax represents the softmax activation function, Represents the normalized matching score vector.

9. The method for sequential recommendation based on multi-behavior dynamic graph according to claim 1, characterized in that: The formula for calculating the total loss of the model is: Among them, LOSS represents the total loss of the model, S represents the sequence set; y uv Indicates whether the next interaction of user u is item v, is the normalized matching score between user u and item v, θ represents all model parameters, ‖·‖2 represents the L2 norm; λ is the regularization strength control coefficient.

Images