A Sequential Recommendation Method Based on Multi-Behavior Dynamic Graph

Abstract

The present invention belongs to the technical field of sequence recommendation, and specifically relates to a sequence recommendation method based on a multi-behavior dynamic graph, including: constructing a multi-behavior dynamic graph according to the multi-behavior interaction sequence of a user and dividing it into multiple specific behavior subgraphs; processing the specific behavior subgraphs to obtain node embedding representations and relationship embedding representations; processing the relationship embedding representations to obtain a general relevance embedding representation; fusing the temporal information under specific behaviors into the nodes of the specific behavior subgraphs to obtain a fused user node embedding representation and a fused item node embedding representation; fusing the general relevance embedding representation and the fused user node embedding representation to obtain a user personalized behavior relevance embedding representation; updating the fused item node embedding representation and aggregating the update result and the fused user node embedding representation to obtain the final two node embedding representations; using a feedforward neural network to process the final two node embedding representations to obtain an item recommendation result; the present invention improves the accuracy of recommendation.

Description

Technical Field

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

[0002] Traditional recommendation systems often use static methods to analyze the interaction between users and commodities. This mode mainly captures the general preferences of users. However, in real-world applications, users' behaviors are continuously changing, and their interaction patterns, preferences, and the popularity of commodities will change over time. To cope with this dynamics, sequence recommendation systems have been introduced.

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

[0004] Given the significant practical value of sequence recommendation systems, the research on sequence recommendation problems is booming, and various innovative solutions have emerged one after another. In this field, some methods based on the idea of Markov chains have been introduced. However, under the assumption of emphasizing the independence of each behavior, the independent combination of past components limits the prediction accuracy. To overcome this limitation, deep learning-based methods have emerged, and 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 represent the dynamic interaction relationship between users and items and learn effective user preference representations. Numerous experiments have confirmed the advantages of these methods, but there are still the following problems:

[0005] Current sequence recommendation methods generally focus on the target behavior data of users to form user-item interaction sequences. However, in the real world, users usually interact with items in a multi-behavior serialized manner. For example, on an e-commerce platform, interaction behaviors include clicking, adding to favorites, adding to the shopping cart, and purchasing, etc. The information in multi-behavior interaction data often provides more fine-grained interest preferences of users. 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 deficiencies in the prior art, the present invention proposes a sequence recommendation method based on a multi-behavior dynamic graph. The method includes: obtaining the multi-behavior interaction sequence of the 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 historical behavior data of the user, and extract the multi-behavior interaction sequence from the historical behavior data;

[0009] S2: Construct a multi-behavior dynamic graph according to 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 an initial node embedding representation; project all the initial node embedding representations into the same feature space to obtain an intermediate node embedding representation;

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

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

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

[0014] S7: Fuse the general correlation embedding representation of every two specific behavior subgraphs and the fused user node embedding representation to obtain a personalized behavior correlation embedding representation of every two specific behavior subgraphs of the user;

[0015] S8: According to the personalized behavior correlation embedding representation of every two specific behavior subgraphs of the user, use an attention mechanism to process the fused item node embedding representation to obtain an updated fused item node embedding representation;

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

[0017] S10: Input the final item node embedding representation and the final user node embedding representation into a feed-forward neural network for processing to obtain a matching score;

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

[0019] Further, 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 the multi-behavior dynamic graph, U represents the user set, V represents the item set, B represents the behavior set, b represents the type of interaction behavior, t represents the timestamp when the interaction occurs, o v,u represents the position of item v among all items interacted by user u, o u,v represents the position of user u among all users who have interacted with item v;

[0022] The multi-behavior dynamic graph is divided into multiple specific behavior subgraphs, which is represented as:

[0023]

[0024] Among them, G b represents the subgraph that only contains behavior b, U b represents the set of users who have interacted with the item with behavior b, V b represents the set of items that have interacted with the user with behavior b, t represents the timestamp when the interaction occurs, represents the position of item v among all items interacted by user u with behavior b, represents the position of user u among all users who have interacted with item v with behavior b.

[0025] Further, the formula for obtaining the relational embedding representation of the 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 the set of intermediate user node representations in the specific behavior subgraph G b , H vFor the specific behavior subgraph G b The intermediate item nodes in it represent sets; h b represents the relational embedding representation of the 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 general correlation embedding representation of every two specific behavior subgraphs is:

[0031]

[0032] where represents the general correlation embedding representation of the i-th specific behavior subgraph and the j-th specific behavior subgraph, represents the relational 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 item node embedding representation under the specific behavior subgraph includes:

[0034] Fuse the relational embedding representation of the specific behavior subgraph, the intermediate node embedding representation of the user, and the intermediate node embedding representation of the item to obtain the first attention score; perform normalization processing on the first attention score to obtain the first attention weight; calculate the fused user node embedding representation according to the first attention weight and the intermediate node embedding representation of the item;

[0035] Fuse the relational embedding representation of the specific behavior subgraph, the intermediate node embedding representation of the item, and the intermediate node embedding representation of the user to obtain the second attention score; perform normalization processing on the second attention score to obtain the second attention weight; calculate the fused item node embedding representation according to 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 every two specific behavior subgraphs of the user is:

[0037]

[0038] where represents the personalized behavior correlation rate between the i-th specific behavior subgraph and the j-th specific behavior subgraph of user u; represents the general correlation embedding representation of the i-th specific behavior subgraph and the j-th specific behavior subgraph; It represents the personalized behavior correlation embedding representation of the i-th and j-th specific behavior subgraphs of user u; σ represents the activation function, and W6 and W7 are the seventh and eighth parameter matrices respectively.

[0039] Furthermore, the formula for obtaining the updated fused item node embedding representation is:

[0040]

[0041] where q i represents the query vector of the i-th item interacting with user u, represents b (u,i) the fused item node embedding representation of the i-th item interacting with user u under the b behavior subgraph, represents the b of user u (u,i) behavior subgraph and b (u,j) the personalized behavior correlation embedding representation of the behavior subgraph, k j represents the key vector of the j-th item interacting with user u, represents b (u,j) the fused item node embedding representation of the j-th item interacting with user u under the b behavior subgraph, v j represents the value vector of the j-th item interacting with user u, A[i, j] represents the attention weight of the j-th item interacting with user u to the i-th item interacting with user u, softmax represents the softmax activation function, N u represents the set of items that have interacted with the user, represents b (u,i) the updated fused item node embedding representation of the i-th item interacting with user u under the b behavior subgraph.

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

[0043]

[0044] where score uv represents 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 item 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, and S represents the sequence set; y uv represents whether the next interaction of user u is item v, is the normalized matching score between user u and item v, θ represents all parameters of the model, ‖·‖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 temporal information of the multi-behavior interaction sequence of user items, uses a multi-behavior dynamic graph to store interaction sequence data, and uses methods such as the attention mechanism to learn and update user embeddings, item embeddings, and behavior embeddings. Then, the final user representation and the candidate item to be predicted are fed into a feed-forward neural network to calculate the matching score. According to the matching score, a more accurate recommended object can be provided for the user, improving the accuracy of the recommendation. BRIEF DESCRIPTION OF THE DRAWINGS

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

[0050] Figure 2 is the training framework diagram of the item recommendation model in the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0051] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

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

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

[0055] S1: Obtain the historical behavior data of the user, and extract the multi-behavior interaction sequence from the historical behavior data.

[0056] Obtain the historical behavior data of users, which contains various behaviors of different users towards different items in the past. For example, in product recommendation, the interaction behavior data such as clicks, collections, and purchases of different products by each user; extract multi-behavior interaction sequences from the historical behavior data. The multi-behavior interaction sequence {u1, v1 click, t1} represents that user u1 performed a click interaction behavior on product v1 at time t1.

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

[0058] S2: Construct a multi-behavior dynamic graph according to the multi-behavior interaction sequence, and divide the multi-behavior dynamic graph into multiple specific behavior subgraphs.

[0059] Obtain the multi-behavior dynamic graph according to the multi-behavior interaction sequence, which is expressed as:

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

[0061] where G represents the multi-behavior dynamic graph, U represents the set of users, V represents the set of items, B represents the set of behaviors, b represents the type of interaction behavior, t represents the timestamp when the interaction occurs, o v,u represents the position of item v among all the items interacted by user u, o u,v represents the position of user u among all the users who have interacted with item v.

[0062] Divide the multi-behavior dynamic graph into multiple specific behavior subgraphs. Take the edges and their connected nodes in the multi-behavior dynamic graph G that only contain the specific behavior b. The division process is expressed as:

[0063]

[0064] where G b represents the subgraph that only contains behavior b, U b represents the set of users who have interacted with items with behavior b, V b represents the set of items that have interacted with users with behavior b, t represents the timestamp when the interaction occurs, represents the position of item v among all the items interacted by user u with behavior b, represents the position of user u among all the users who have interacted with item v with behavior b.

[0065] 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.

[0066] Initialize the node embedding representation of the specific behavior subgraph. For the nodes in the behavior subgraph, use the random walk method to obtain the initial embedding representation of the nodes.

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

[0068] h u = W0x u

[0069] h v = W1x v

[0070] where, h u is the intermediate node embedding representation of the user obtained after projection, and h v is the intermediate node embedding representation of the item obtained after projection; x u represents the embedding representation of the user node, and x v represents the embedding representation of the item node; d i is the dimension after the projection of the node representation, d u is the initial dimension of the user node representation, and d v is the initial dimension of the item node representation; 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 the specific behavior subgraph to obtain the relationship embedding representation of the specific behavior subgraph.

[0072] The multi-behavior encoder generates a relationship representation from the nodes on the relationship-specific subgraph, and finally obtains the relationship embedding representation of the specific behavior subgraph; first, aggregate the user representation and the item representation in the behavior subgraph G b respectively, which is expressed as:

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

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

[0075] where, H u is the specific behavior subgraph G bThe intermediate user nodes in represent a set, H v is the intermediate item node in the specific behavior subgraph G b represents a set; M 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] Subsequently, a relational embedding representation of the specific behavior subgraph is generated, denoted as:

[0077] h b = σ(W2[M u ‖M v +b0)

[0078] where σ represents the activation function; ‖ represents the vector concatenation operation; b0 represents the first bias; and W2 represents the third parameter matrix.

[0079] S5: Use the general behavior correlation encoder to process the relational embedding representations of any two specific behavior subgraphs to obtain the general correlation embedding representations of every two specific behavior subgraphs.

[0080] Use the general behavior correlation encoder to learn any set of relation pairs to obtain the general correlation embedding representations of any two relations:

[0081]

[0082] where represents the general correlation embedding representation of the i-th specific behavior subgraph and the j-th specific behavior subgraph; is a set of behavior relation pairs, represents the relational 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: Incorporate 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.

[0084] The interaction behaviors of the user with the item at different times actually reflect the dynamic changes in the user's interest preferences. Therefore, use the attention mechanism to incorporate the temporal information under the specific behavior into the user nodes to obtain the new user node representation, that is, the fused user node embedding representation under the specific behavior subgraph:

[0085]

[0086] where denotes 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 the position encoding vector directly related to calculating the attention score; is the first attention score of item v for user u under behavior b; is the attention weight obtained after normalization; is the position encoding vector directly related to information propagation; W4 is the fifth parameter matrix.

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

[0088]

[0089] where, denotes the fused item node embedding representation of item v under the feature behavior subgraph, is the set of all users who have interacted with item v under behavior b; is the position encoding vector directly related to calculating the attention score; is the second attention score of user u for item v under behavior b; is the attention weight obtained after normalization; is the position encoding vector directly related to information propagation; W5 is the sixth parameter matrix.

[0090] S7: Fuse the general relevance embedding representation of every two specific behavior subgraphs and the fused user node embedding representation to obtain the personalized behavior relevance embedding representation of every two specific behavior subgraphs of the user.

[0091] Fuse the general relevance embedding representation of every two specific behavior subgraphs and the fused user node embedding representation:

[0092]

[0093] where, denotes the personalized behavior correlation rate between the i-th and j-th specific behavior subgraphs of the user; It represents the personalized behavior correlation embedding representation of the i-th and j-th specific behavior subgraphs of the user; W6 and W7 are the seventh and eighth parameter matrices respectively.

[0094] S8: According to the personalized behavior correlation embedding representations of every two specific behavior subgraphs of the user, the attention mechanism is used to process the fused item node embedding representation to obtain the updated fused item node embedding representation.

[0095] The attention mechanism is used to fuse the personalized behavior correlation and the temporal information into the item representation to update the item representation:

[0096]

[0097] where q i represents the query vector of the i-th item interacting with user u, represents b (u,i) the fused item node embedding representation of the i-th item interacting with user u under the b behavior subgraph, represents the b of the user (u,i) behavior subgraph and b (u,j) the personalized behavior correlation embedding representation of the behavior subgraph, k j represents the key vector of the j-th item interacting with user u, represents b (u,j) the fused item node embedding representation of the j-th item interacting with user u under the b behavior subgraph, v j represents the value vector of the j-th item interacting with user u, A[i, j] represents the attention weight of the j-th item interacting with user u to the i-th item interacting with user u, softmax represents the softmax activation function, N u represents the set of items that have interacted with the user, represents b (u,i) the updated fused item node embedding representation of the i-th item interacting with user u under the b behavior subgraph.

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

[0099] Final item node embedding representation:

[0100]

[0101] Final user node embedding representation:

[0102]

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

[0104] Input the embedded representations of the final project node and the final user node into a feedforward neural network for processing:

[0105]

[0106] where score uv represents the matching score between user u and project 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 loss of the model according to the matching score, and adjust the model parameters according to the total loss of the model. When the loss function converges, the training of the model is completed.

[0108] The present invention uses binary cross-entropy as the loss function of the recommendation task, that is, the total loss of the model, and calculates the total loss of the model:

[0109]

[0110] where LOSS represents the total loss of the model, S represents the sequence set; y uv is the true label, indicating whether the next interaction of the user is project v. Take 1 when it is v, otherwise take 0; is the normalized matching score between user u and project v, θ represents all parameters of the model, ‖·‖2 represents the L2 norm; λ is the regularization strength control coefficient, 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 training of the model, and obtain a trained project recommendation model. Obtain the multi-behavior interaction sequence of the user to be recommended; input the multi-behavior interaction sequence into the trained project recommendation model, and the best project recommendation result can be obtained.

[0112] The above embodiments further elaborate on the purpose, technical solution, and advantages of the present invention. It should be understood that the above embodiments are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made to the present invention within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A sequence recommendation method based on a multi-behavior dynamic graph, characterized in that, Including: 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; The training process of the item recommendation model includes: S1: Obtain the historical behavior data of the user, and extract the multi-behavior interaction sequence from the historical behavior data; S2: Construct a multi-behavior dynamic graph according to the multi-behavior interaction sequence, and divide the multi-behavior dynamic graph into multiple specific behavior sub-graphs; 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 the user set, V represents the item set, B represents the behavior set, b represents the type of interaction behavior, t represents the timestamp when the interaction occurs, and o v,u represents the position of item v among all items interacted by user u, and o u,v represents the position of user u among all users who have interacted with item v; Dividing the multi-behavior dynamic graph into multiple specific behavior sub-graphs is represented as: Among them, G b represents the subgraph that only contains behavior b, U b represents the set of users who have interacted with the project with behavior b, V b represents the set of projects that have interacted with users with behavior b, t represents the timestamp when the interaction occurred, represents the position of project v among all projects that user u has interacted with using behavior b, represents the position of user u among all users who have interacted with project v using behavior b; S3: Initialize the node embedding representation of the specific behavior sub-graph to obtain the initial node embedding representation; Project all the initial node embedding representations into the same feature space to obtain the intermediate user node embedding representation and the intermediate item node embedding representation; S4: Use a multi-behavior encoder to process the intermediate user node embedding representation and the intermediate item node embedding representation of the specific behavior sub-graph to obtain the relationship embedding representation of the specific behavior sub-graph; S5: Use a general behavior correlation encoder to process the relationship embedding representations of any two specific behavior sub-graphs to obtain the general correlation embedding representation of every two specific behavior sub-graphs; S6: Incorporate the temporal information under the specific behavior into the nodes of the specific behavior sub-graph to obtain the fused user node embedding representation and the fused item node embedding representation under the specific behavior sub-graph; S7: Fuse the general correlation embedding representation of every two specific behavior sub-graphs and the fused user node embedding representation to obtain the personalized behavior correlation embedding representation of every two specific behavior sub-graphs of the user; S8: According to the personalized behavior correlation embedding representation of every two specific behavior sub-graphs of the user, use the attention mechanism to process the fused item node embedding representation to obtain the updated fused item node embedding representation; S9: Aggregate the updated fused item node embedding representations under each specific behavior sub-graph to obtain the final item node embedding representation; Aggregate the fused user node embedding representations under each specific behavior sub-graph to obtain the final user node embedding representation; S10: Input the final item node embedding representation and the final user node embedding representation into a feed-forward neural network for processing to obtain the matching score; S11: Calculate the total loss of the model according to the matching score, and adjust the model parameters according to the total loss of the model. When the loss function converges, the training of the model is completed.

2. The sequence recommendation method based on a multi-behavior dynamic graph according to claim 1, wherein The formula for obtaining the relationship embedding representation of the specific behavior sub-graph 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 the set of intermediate user nodes in the specific behavior subgraph G b ; H v is the set of intermediate item nodes in the specific behavior subgraph G b ; h b represents the relational embedding representation of the specific behavior subgraph; σ represents the activation function; || represents the vector concatenation operation; M u represents the aggregated user representation, and M v represents the aggregated item representation; b0 represents the first bias; W2 represents the third parameter matrix.

3. A sequence recommendation method based on a multi-behavior dynamic graph according to claim 1, characterized in that The formula for obtaining the general correlation embedding representation of every two specific behavior sub-graphs is: Among them, represents the general correlation embedding representation of the i-th specific behavior sub-graph and the j-th specific behavior sub-graph, represents the relational embedding representation of the i-th specific behavior sub-graph, represents the relational embedding representation of the j-th specific behavior sub-graph, b1 represents the second bias, σ represents the activation function; || represents the vector concatenation operation, and W3 represents the fourth parameter matrix.

4. The sequence recommendation method based on a 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 the specific behavior sub-graph includes: Fuse the relationship embedding representation of the specific behavior sub-graph, the intermediate node embedding representation of the user, and the intermediate node embedding representation of the item to obtain the first attention score; Normalize the first attention score to obtain the first attention weight; Calculate the fused user node embedding representation according to the first attention weight and the intermediate node embedding representation of the item; It is represented by the formula: Among them, 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 related to the position encoding vector for calculating the attention score directly; is the first attention score of item v for user u under behavior b; is the attention weight obtained after normalization; represents the position of item v among all items that user u has interacted with under behavior b; is related to the position encoding vector for information propagation directly related to; W4 is the fifth parameter matrix; h u is the intermediate node embedding representation of the user, h v is the intermediate node embedding representation of the item; σ represents the activation function; Fuse the relationship embedding representation of specific behavior subgraphs, the intermediate node embedding representation of items, and the intermediate node embedding representation of users to obtain the second attention score; perform normalization processing on the second attention score to obtain the second attention weight; calculate the fused item node embedding representation according to the second attention weight and the intermediate node embedding representation of users; which is expressed by the formula: Among them, represents the fused item node embedding representation of item v under the characteristic behavior subgraph; is the set of all users who have interacted with item v under behavior b; is related to the position encoding vector for calculating the attention score directly related; is the second attention score of user u for item v under behavior b; is the attention weight obtained after normalization; represents the position of user u among all users who have interacted with item v with behavior b; is related to the position encoding vector for information propagation directly related; W5 is the sixth parameter matrix.

5. A sequence recommendation method based on a multi-behavior dynamic graph according to claim 1, characterized in that The formula for obtaining the personalized behavior correlation embedding representation of every two specific behavior subgraphs of users is: Among them, represents the personalized behavior correlation rate between the i-th and j-th specific behavior subgraphs of user u; represents the general correlation embedding representation between the i-th and j-th specific behavior subgraphs; represents the personalized behavior correlation embedding representation between the i-th and j-th specific behavior subgraphs of user u; σ represents the activation function, and W6 and W7 are the seventh and eighth parameter matrices respectively.

6. The sequence recommendation method based on a multi-behavior dynamic graph according to claim 1, wherein The formula for obtaining the updated fused item node embedding representation is: where q i represents the query vector of the i-th item interacted with user u, represents b (u,i) is the fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph, represents b of user u (u,i) behavior subgraph and b (u,j) personalized behavior relevance embedding representation of the behavior subgraph, k j represents the key vector of the j-th item interacted with user u, represents b (u,j) is the fused item node embedding representation of the j-th item interacted with user u under the behavior subgraph, v j represents the j-th item interacted with user u, A[i, j] represents the attention weight of the j-th item interacted with user u to the i-th item interacted with user u, softmax represents the softmax activation function, N u represents the set of items interacted with the user, represents b (u,i) is the updated fused item node embedding representation of the i-th item interacted with user u under the behavior subgraph.

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

8. A sequence recommendation method based on a 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 set of sequences; 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 parameters of the model, ‖·‖2 represents the L2 norm; λ is the regularization strength control coefficient.

Images