A Graph Collaborative Filtering Recommendation Method Based on Representation Disentanglement

Abstract

The present invention relates to a graph collaborative filtering recommendation method based on representation decoupling, belonging to the field of computer technology. The method includes the following steps: S1: Obtain a data set; S2: Perform data preprocessing; S3: Construct a user-item bipartite graph from the data; S4: Construct a graph convolutional module to generate user-item embeddings by combining user features, item features, and graph structure; S5: Construct a top-K generator to form new user-item bipartite graphs, user-user bipartite graphs, and item-item bipartite graphs; S6: The graph decoupling layer decouples the pre-trained user-item embeddings to generate the final user-item embeddings, and obtains the final recommendation result through inner product. The present invention solves the problems of unclear embedding semantics caused by user-item interaction noise and the graph convolutional layer in the graph collaborative filtering-based recommendation algorithm, so that the recommendation performance is improved on the original graph collaborative filtering model.

Description

Technical Field

[0001] The present invention belongs to the technical field of computers and relates to a graph collaborative filtering recommendation method based on representation decoupling. Background Art

[0002] With the proposal of graph learning (GL) methods, especially graph neural networks (GNNs) in graph learning, which have achieved great success in many graph tasks such as complex relationship extraction and link prediction. Since most data in recommendation systems has the characteristics of a graph structure, it is a natural approach to apply graph learning methods to the recommendation field. Therefore, graph learning methods have gradually become an emerging recommendation paradigm widely studied in the recommendation field, namely, recommendation systems based on graph learning. When the interaction information between users and items is constructed in the form of a graph structure and combined with graph learning methods mainly represented by graph neural networks, it is possible to capture, learn, and simulate the high-order and complex relationships between users and items, and more effectively learn the long-term interest preferences of users and the characteristic attributes of items to improve the recommendation performance of the recommendation system.

[0003] With the wide attention paid to the field of interpretable machine learning, the interpretability of recommendation systems has also become a research hotspot. Due to the characteristics of the graph convolutional network itself, the embedding representations generated by models based on graph collaborative filtering cannot correspond one-to-one between the representations and the semantics they represent. Therefore, some researchers have introduced semantic decoupling methods in the field of natural language processing into recommendation models based on graph collaborative filtering, enabling the representations generated by the models to correspond one-to-one with their semantics through decoupling, thereby increasing the interpretability of the representations. However, the current decoupling-based models are still limited by problems such as sparse user-item interaction data and noise, and there is still much room for performance improvement. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a graph collaborative filtering recommendation method based on representation decoupling, which solves the technical problem that the decoupling model cannot maximize the performance due to sparse user-item interaction data and noise when decoupling the semantics in pictures.

[0005] To achieve the above purpose, the present invention provides the following technical solutions:

[0006] A graph collaborative filtering recommendation method based on representation decoupling, the method comprising the following steps:

[0007] S1: Obtain a data set, the data set comprising user information U and item information I;

[0008] S2: Preprocess the text data in the dataset obtained in S1, divide the preprocessed dataset into a training set train.txt and a test set test.txt. The training set is used for model training, and the test set is used for model performance testing. Perform the following operations on the two datasets respectively;

[0009] S3: Construct a user-item interaction matrix based on the dataset, construct a user-item interaction bipartite graph based on the user-item interaction matrix, and construct an adjacency matrix based on the user-item interaction bipartite graph; generate user-item embeddings by combining the constructed user-item interaction matrix and adjacency matrix. denotes the initial embedding of user u, denotes the initial embedding of item i;

[0010] S4: Input the initial embeddings and in S3 into the first graph convolutional module for graph convolutional operations, and learn the embeddings of users and items by iteratively aggregating the features of adjacent nodes in the bipartite graph.

[0011] S5: Construct a top-K generator, and use the user and item embeddings output by the first graph convolutional module to reconstruct the user-item interaction bipartite graph, user-user bipartite graph, and item-item bipartite graph;

[0012] S6: Generate a new adjacency matrix according to the three newly reconstructed interaction graphs in S5, and input the generated new matrix and the user and item embeddings obtained in S4 into the second graph convolutional module for graph convolutional operations to obtain new user and item embeddings;

[0013] S7: Input the new user and item embeddings obtained in S6 and the user-item interaction bipartite graph newly constructed in S5 into the graph disentanglement layer for graph representation disentanglement operations to obtain the final representations of users and items. The model prediction is defined as the inner product of the final representations of users and items, and the final recommendation result is obtained through the inner product.

[0014] Furthermore, in S3, the user-item interaction matrix is denoted as: R ∈ R^(|N|×|M|), where N and M represent the numbers of users and items respectively; r ui ∈ R is a non-zero entry, indicating that user u ∈ U has interacted with item i ∈ I; if not, the entry is zero;

[0015] The user-item interaction bipartite graph is denoted as: G = (W, E), where W is the node set containing user nodes and item nodes, and E is the set of edges; for non-zero r ui , there is an edge between user u and item i; r uiIt is an element in the user-project interaction matrix R, where u represents the u-th row and i represents the i-th column;

[0016] Adjacency matrix, denoted as: A ∈ R (|N|+|M|)×(|M|+|N|) .

[0017] Furthermore, in the step S4, the graph convolution operation is denoted as:

[0018]

[0019]

[0020] Among them, and respectively represent the embeddings of user u and item i after k-layer propagation; N u is the set of items associated with user u;

[0021] Iteratively aggregate the features of adjacent nodes in the bipartite graph to learn the embeddings of users and items, denoted as:

[0022]

[0023] Among them, AGG is an aggregation function, which is the core of graph convolution and is used to obtain the embeddings of the target node and its neighbor nodes at the k-th layer. is the symmetric normalization term, and the matrix L is the Laplacian matrix of the adjacency matrix A. D is the degree matrix of matrix A;

[0024] After K-layer graph convolution, the final embedding of a user-item is the weighted sum of the embeddings obtained at each layer in the graph convolution module, denoted as:

[0025]

[0026]

[0027] Among them, α k ≥ 0 is a hyperparameter assigned to the k-th layer, which controls the proportion of the embedding output by the current layer in the final embedding.

[0028] Furthermore, the step S5 is specifically as follows: construct a top-K generator, take the inner product of the embeddings of user-items obtained from the first graph convolution module, and according to the inner product results, obtain the top K items that the target user is most interested in, the top K users that are most similar to the target user, and the top K items that are most similar to the target item; construct a new user-item bipartite graph, user-user bipartite graph, and item-item bipartite graph in the top-K generator according to the obtained results, denoted as:

[0029]

[0030]

[0031]

[0032] Among them, E U and E I are the user embedding matrix and the item embedding matrix obtained by the first graph convolution module, respectively.

[0033] Furthermore, the S7 inputs the new user and item embeddings obtained in the S6 and the newly constructed user-item interaction bipartite graph in the S5 into the graph disentanglement layer for graph representation disentanglement operation to obtain the final representations of the user and the item, which specifically includes the following steps:

[0034] S51: Divide the embeddings input into the graph disentanglement layer into h blocks, and make each block associated with a potential intention, expressed as:

[0035] e u =(e u1 , e u2 , …, e uH ) (9)

[0036] e i =(e i1 , e i2 , …, e iH ) (10)

[0037] Among them, e u and e i represent the user embedding and the item embedding of the target user u and item i obtained in the S6 respectively; e uh represents the part related to the intention h in the embedding e u .

[0038] S52: Define a set of score matrices Any entry S h(u,i) in the matrix represents the interaction generated by the user u for the item i under the action of the intention h; from the perspective of the interaction between the user and the item, then S (u,i) =(S 1(u,i) , S 2(u,i) , …, S H(u,i) ), and S (u,i) concentrates all the intentions that generate the current interaction into a set, representing the proportion of the contribution of the intention h in the occurred interactions;

[0039] S53: Propagate the embeddings divided into h blocks according to the intention-aware bipartite graph corresponding to the score matrix in the S52, expressed as:

[0040]

[0041] Among them, represents the multi-hop neighbors of the user, represents the proportion of the intention h in the interaction behavior generated by the user u for i after iterating to the t-th layer, that is, after t rounds of iteration, represents any element in formula (10);

[0042] S54: Aggregate the user intention perception embeddings obtained in each layer, that is, obtain the final user intention perception embedding, which is expressed as:

[0043]

[0044] Then aggregate the intention-based perception embeddings, that is, obtain the final user representation, which is expressed as:

[0045]

[0046] Repeat the above S51 - S54, and also obtain the final representation of the item

[0047] Furthermore, in the above S7, the model prediction is defined as the inner product of the final representations of the user and the item, and the final recommendation result is obtained through the inner product, specifically:

[0048] The learned user representation and the item representation Calculate the preference score of the user u for the item i through the inner product, and then the final recommendation result can be obtained, which is expressed as:

[0049]

[0050] Use the loss function BPR as the target optimization function, which is expressed as:

[0051]

[0052] Among them, O is the training data set, i represents the item that has an interaction with the user u, j represents the item that has no interaction with the user u, σ(·) is the sigmoid function, λ is the coefficient to control L2 regularization, and Θ is the set of all trainable parameters in the model, represents the preference score of the user for the item that has no explicit interaction with it. The beneficial effects of the present invention are as follows:

[0053] First, a graph collaborative filtering recommendation method based on representation decoupling provided by the present invention preprocesses the original interaction matrix of user items, and regenerates a new user-item interaction bipartite graph by screening out items of interest to the target user, users with similar preferences to the target user, and items similar to the target item, and learns node embeddings based on the new interaction graph to alleviate the data sparsity and noise problems encountered by the current model.

[0054] Second, the present invention proposes to construct a top-K generator, which reconstructs the user-item interaction bipartite graph, the user-user bipartite graph, and the item-item bipartite graph by using the user and item representations output by the first graph convolutional module, thereby effectively alleviating the noise and sparsity problems existing in the original interaction graph.

[0055] Third, the present invention proposes to construct a graph decoupling layer to decouple the representations according to different intentions, so as to correspond each potential intention to a certain part of the representation, thereby achieving better recommendation results.

[0056] Other advantages, objectives and features of the present invention will be described to some extent in the subsequent description, and to some extent, will be obvious to those skilled in the art based on the study of the following text, or can be learned from the practice of the present invention. The objectives and other advantages of the present invention can be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be described in preferred detail below with reference to the drawings, where:

[0058] Figure 1 is a flowchart of the graph collaborative filtering recommendation algorithm model for representation decoupling of the present invention;

[0059] Figure 2 is a schematic diagram of the graph convolutional module of the present invention;

[0060] Figure 3 is a flowchart of the representation decoupling of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0061] The following specific examples illustrate the embodiments of the present invention. Those skilled in the art can easily understand other advantages and effects of the present invention from the content disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and various details in this specification can also be modified or changed based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the diagrams provided in the following embodiments only illustrate the basic concept of the present invention schematically, and the following embodiments and the features in the embodiments can be combined with each other without conflict.

[0062] Among them, the accompanying drawings are only for illustrative purposes, showing only schematic diagrams, not physical diagrams, and should not be construed as limiting the present invention; in order to better illustrate the embodiments of the present invention, some components in the accompanying drawings will be omitted, enlarged or reduced, which does not represent the size of the actual product; for those skilled in the art, it is understandable that some well-known structures and their descriptions in the accompanying drawings may be omitted.

[0063] In the accompanying drawings of the embodiments of the present invention, the same or similar reference numerals correspond to the same or similar components; in the description of the present invention, it should be understood that if there are terms such as "upper", "lower", "left", "right", "front", "rear", etc. indicating the orientation or positional relationship, they are based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, the terms describing the positional relationship in the accompanying drawings are only for illustrative purposes and should not be construed as limiting the present invention. For those of ordinary skill in the art, the specific meanings of the above terms can be understood according to specific circumstances.

[0064] Please refer to Figures 1 to 3 , which is a graph collaborative filtering recommendation method based on representation decoupling.

[0065] What the present invention discloses is a recommendation algorithm based on graph collaborative filtering, and the detailed steps are as follows:

[0066] Step 1: Obtain a data set, including user information and item information. The data used in the present invention is the Amazon-Book data set, which contains 52,643 pieces of user information and 91,599 pieces of item information;

[0067] Step 2: Preprocess the text data and divide the data into train.txt and test.txt for separate processing;

[0068] Step 3: Let \(R\in\mathbb{R}^{(|N|\times|M|)}\) represent the user-item interaction matrix, where \(N\) and \(M\) represent the number of users and items respectively. The non-zero entry \(r_{ui}\in\mathbb{R}\) indicates that user \(u\in U\) has interacted with item \(i\in I\); otherwise, the entry is zero; according to the interaction matrix, construct a bipartite graph \(G=(W, E)\) of users and items, where the node set \(W\) contains user nodes and item nodes, and \(E\) is the set of edges; for non-zero \(r\) ui , there is an edge between user \(u\) and item \(i\). Construct the adjacency matrix \(A\in\mathbb{R}\) (|N|+|M|)×(|M|+|N|) . Represents the initial embedding of user \(u\), Denote the initial embedding of item i. The above information is used as the input of the graph convolutional layer, and the representations of users and items are learned by iteratively aggregating the features of adjacent nodes in the bipartite graph. The graph convolution operation is as follows:

[0069]

[0070]

[0071] Step 4: The main idea of the graph convolution module is to update the node representation by aggregating the features of each node in the graph. To achieve this goal, graph convolution needs to be performed iteratively, that is, aggregate the embedding representations of the neighbors around the target node to this node to generate a new embedding representation of the target node. This way of aggregating neighbor node information can be expressed by the formula:

[0072]

[0073] In the formula: and represent the embeddings of user u and item i after k-layer propagation respectively; N u is the set of items associated with user u; AGG is an aggregation function, which is the core of graph convolution and is used to obtain the embedding representations of the target node and its neighbor nodes in the k-th layer. is a symmetric normalization term, the matrix L is the Laplacian matrix of the adjacency matrix A, D is the degree matrix of matrix A;

[0074] Step 5: After K-layer graph convolution, the final embedding of a user is the weighted sum of the embeddings obtained at each layer in the graph convolution module:

[0075]

[0076]

[0077] where α k ≥ 0 is a hyperparameter assigned to the k-th layer, which is used to control the proportion of the embedding output by the current layer in the final embedding. In actual application scenarios, the interactions between users and items often have noise due to users' misclicks, browsing, observation and other behaviors. At the same time, many items belong to long-tail items, which often only interact with one or two users, resulting in sparse interaction data. And through graph convolution operations, the influence of noise and popular items will be further amplified, eventually leading to a series of problems that degrade performance such as suboptimal recommendations;

[0078] Step 6: To solve the above problems, the present invention proposes a top-K generator. By using the user and item representations output by the first graph convolutional module, the user-item interaction bipartite graph, user-user bipartite graph, and item-item bipartite graph are reconstructed, thereby effectively alleviating the noise and sparsity problems existing in the original interaction graph. To achieve this goal, the inner product of the user-item embeddings obtained by the graph convolutional module is calculated. According to the inner product results, the top K items that the target user is most interested in, the top K users that are most similar to the target user, and the top K items that are most similar to the target item are obtained; then, based on the above results, a new user-item bipartite graph, user-user bipartite graph, and item-item bipartite graph are constructed. As follows:

[0079]

[0080]

[0081]

[0082] where, E U and E I are the user embedding matrix and item embedding matrix obtained by the graph convolutional module respectively. After generating the new interaction graph through Step 6, a new adjacency matrix is generated using the interaction graph. At the same time, the user and item representations obtained in Step 5 and the new adjacency matrix are input into the second graph convolutional module. Through graph convolutional operations, new user and item representations are obtained. The new representations are more robust to noise signals and can represent user preferences more comprehensively and meticulously. There are often different intentions behind a user's choice to interact with an item. It is extremely necessary to distinguish these multiple intentions and map them one by one to the representations. Especially in a recommendation model based on a graph convolutional model framework, when the target node aggregates the information of the surrounding neighborhood, the neighborhood nodes are often regarded as a whole, and it is not possible to well distinguish the intention behind the connection between different neighborhood nodes and the target node. To solve this problem, the present invention proposes to construct a graph decoupling layer to decouple the representations according to different intentions, so as to map each potential intention to a certain part of the representation, thereby achieving better recommendation results

[0083] Specifically, the user and item representations output by the second graph convolutional module and the user-item interaction graph obtained in Step 6 are input into the graph disentanglement layer. The representations are divided into h blocks, and each block is associated with a potential intention. As follows

[0084] e u =(e u1 ,e u2 ,…,e uH ) (9)

[0085] e i =(ei1 , e i2 , …, e iH ) (10)

[0086] Among them, e u and e i respectively represent the user embedding and item embedding obtained by the target user u and item i in the second graph convolutional module; e uh represents the part related to the intention h in the embedding e u ;

[0087] Define a set of score matrices Any entry S in the matrix h(u,i) represents the interaction generated by user u on item i under the action of intention h; at the same time, from the perspective of the interaction between the user and the item, there is S (u,i) =(S 1(u,i) , S 2(u,i) , …, S H(u,i) ), indicating the proportion of the contribution made by intention h in the existing interactions;

[0088] The embedding divided into h blocks propagates information according to the intention-aware bipartite graph corresponding to the above score matrix, that is:

[0089]

[0090] In the formula: represents the multi-hop neighbor of the user, represents the proportion of intention h in the interaction behavior generated by user u on i after iterating to the t-th layer (i.e., iterating t rounds);

[0091] Aggregate the user intention-aware embeddings obtained in each layer, and the final user intention-aware representation is obtained:

[0092]

[0093] Then aggregate the intention-based perceptual representations to obtain the final user representation:

[0094]

[0095] Similarly, the final representation on the item side can be obtained

[0096] Use the learned user representation and item representation to calculate the preference score of user u for item i through the inner product:

[0097]

[0098] Finally, the BPR (Bayesian Personalized Ranking) loss function is used as the target optimization function, as follows:

[0099]

[0100] Where: O is the training data set, i represents the item that has an interaction with user u, j represents the item that has no interaction with user u, -(·) is the sigmoid function, λ is the coefficient controlling L2 regularization, and Θ is the set of all trainable parameters in the model. represents the preference score of the user for the item with which there is no explicit interaction.

[0101] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the purpose and scope of the present technical solution, and they should all be covered by the scope of the claims of the present invention.

Claims

1. A graph collaborative filtering recommendation method based on representation decoupling, characterized in that: The method includes the following steps: S1: Obtain a dataset, where the dataset contains user information U and item information I; S2: Preprocess the text data in the dataset obtained in S1, and divide the preprocessed dataset into a training set train.txt and a test set test.txt. The training set is used for model training, and the test set is used for model performance testing. The following steps are performed on the two datasets respectively; S3: Construct a user-item interaction matrix based on the dataset, construct a user-item interaction bipartite graph based on the user-item interaction matrix, and construct an adjacency matrix based on the user-item interaction bipartite graph; generate user-item embeddings by combining the constructed user-item interaction matrix and adjacency matrix, represents the initial embedding of user u, represents the initial embedding of item i; S4: Input the initial embedding in S3 and into the first graph convolutional module for graph convolutional operations, and learn the embeddings of users and items by iteratively aggregating the features of adjacent nodes in the bipartite graph; S5: Construct a top-K generator, and use the user and item embeddings output by the first graph convolutional module to reconstruct the user-item interaction bipartite graph, the user-user bipartite graph, and the item-item bipartite graph; S6: According to the user-item interaction bipartite graph, the user-user bipartite graph, and the item-item bipartite graph reconstructed in S5, generate a new adjacency matrix, and input the new adjacency matrix and the user and item embeddings obtained in S4 into the second graph convolutional module for graph convolutional operations to obtain new user and item embeddings; S7: Input the new user and item embeddings obtained in S6 and the user-item interaction bipartite graph newly constructed in S5 into the graph disentanglement layer for graph representation disentanglement operations to obtain the final representations of the user and the item. The model prediction is defined as the inner product of the final representations of the user and the item, and the final recommendation result is obtained through the inner product.

2. The graph collaborative filtering recommendation method based on representation decoupling according to claim 1, characterized in that: In the above S3, the user-item interaction matrix is represented as: \(R\in\mathbb{R}^{|\mathcal{N}|\times|\mathcal{M}|}\), where \(|\mathcal{N}|\) and \(|\mathcal{M}|\) represent the number of users and items respectively; \(r_{ui}\in\mathbb{R}\) is a non-zero entry, indicating that user \(u\in\mathcal{U}\) has interacted with item \(i\in\mathcal{I}\); if there is no interaction, the entry is zero. ui \(\in\mathbb{R}\) is a non-zero entry, indicating that user \(u\in\mathcal{U}\) has interacted with item \(i\in\mathcal{I}\); if there is no interaction, the entry is zero. The user-project interaction bipartite graph is denoted as: G = (W, E), where W is the node set containing user nodes and project nodes, and E is the set of edges; for non-zero r ui , there is an edge between user u and project i; r ui is an element in the user-project interaction matrix R, where u represents the u-th row and i represents the i-th column; Adjacency matrix, expressed as: A ∈ R (|N|+|M|)×(|M|+|N|) .

3. The graph collaborative filtering recommendation method based on representation decoupling according to claim 2, characterized in that: In S4, the graph convolutional operation is expressed as: Among them, and respectively represent the embeddings of user u and item i after propagation in layer k; N u is the set of items associated with user u; Iteratively aggregate the features of adjacent nodes in the bipartite graph to learn the embeddings of the user and the item, which is expressed as: Among them, AGG is an aggregation function, which is the core of graph convolution and is used to obtain the embeddings of the target node and its neighbor nodes at the k-th layer. is a symmetric normalization term, and the matrix L is the Laplacian matrix of the adjacency matrix A. D is the degree matrix of the matrix A. After K layers of graph convolution, the final embedding of a user-item is the weighted sum of the embeddings obtained in each layer of the graph convolutional module, which is expressed as: Among them, α k ≥ 0 is a hyperparameter assigned to the k-th layer, which controls the proportion of the embedding output by the current layer in the final embedding.

4. The graph collaborative filtering recommendation method based on representation decoupling according to claim 3, wherein: S5 is specifically as follows: Construct a top-K generator, perform an inner product on the embeddings of the user-item obtained by the first graph convolutional module, and according to the inner product result, obtain the top K items that the target user is most interested in, the top K users most similar to the target user, and the top K items most similar to the target item; construct a new user-item bipartite graph, a user-user bipartite graph, and an item-item bipartite graph in the top-K generator according to the obtained results, which are respectively expressed as: Among them, E U and E I are the user embedding matrix and the item embedding matrix obtained by the first graph convolution module, respectively.

5. The graph collaborative filtering recommendation method based on representation decoupling according to claim 4, wherein: In S7, input the new user and item embeddings obtained in S6 and the user-item interaction bipartite graph newly constructed in S5 into the graph disentanglement layer for graph representation disentanglement operations to obtain the final representations of the user and the item, which specifically includes the following steps: S51: Divide the embeddings input into the graph disentanglement layer into h blocks, so that each block is associated with a potential intention, which is expressed as: e u = (e u1 , e u2 , …, e uH ) (9) e i = (e i1 , e i2 ,....., e iH ) (10) Among them, e u and e i respectively represent the user embedding and item embedding obtained by the target user u and item i in S6; e uh denotes the part regarding the intention h in the embedding e u ; S52: Define a set of scoring matrices Any entry S in the matrix h(u，i) represents the interaction generated by user u for item i under the action of intention h; from the perspective of the interaction between the user and the item, then S(u, i) = (S 1(u，i) , S 2(u，i) ,..., S H(u，i) ), S (u，i) Aggregates all the intentions that generate the current interaction into a set, representing the proportion of the contribution made by intention h in the interactions that have occurred; S53: Propagate the information of the embeddings divided into h blocks according to the intention-aware bipartite graph corresponding to the score matrix in S52, which is expressed as: Among them, represents the multi-hop neighbors of the user, represents the proportion of intention h in the interaction behavior generated by user u for i after iterating to the t-th layer, that is, after t rounds of iteration, represents any element in formula (10); S54: Aggregate the user intention-aware embeddings obtained in each layer, that is, obtain the final user intention-aware embedding, which is expressed as: Then aggregate the intention-based perception embeddings, that is, obtain the final user representation, which is expressed as: Repeat the above S51 - S54 to also obtain the final characterization of the project 6. The graph collaborative filtering recommendation method based on representation decoupling according to claim 5, wherein: In S7, the model prediction is defined as the inner product of the final representations of the user and the item, and the final recommendation result is obtained through the inner product, which is specifically as follows: User representation after learning and item representation By calculating the preference score of user u for item i through the inner product, the final recommendation result can be obtained, which is expressed as: Use the loss function BPR as the target optimization function, which is expressed as: Among them, \(O\) is the training dataset, \(i\) represents the item that has an interaction with user \(u\), \(j\) represents the item that has no interaction with user \(u\), \(\sigma(\cdot)\) is the sigmoid function, \(\lambda\) is the coefficient controlling L2 regularization, and \(\Theta\) is the set of all trainable parameters in the model. It represents the preference score of the user for the item with which there is no explicit interaction.

Images