A Recommendation Method and System Based on Multi-View Fusion of Latent Interests

Abstract

The present invention discloses a recommendation method and system based on multi-view fusion of latent interests, belonging to the technical field of computer recommendation systems. The present invention explores the influence of various factors on users and commodities by constructing multiple views, generates subgraphs according to potential user interest themes, and in different views, makes the representations of users and commodities maintain their own characteristics by adding personalized features, effectively alleviating the over-smoothing problem of traditional methods. The classic LightGCN model is improved with a residual network structure, and a VS-LightGCN module is designed to maintain the model's continuous recognition of individualized features. The dynamic multi-view fusion module based on the principal component control mechanism (PCCM) can effectively fuse the generated multi-interest views and improve the recommendation effect of the model.

Description

Technical Field

[0001] The present invention belongs to the technical field of computer recommendation systems, and particularly relates to a recommendation method and system based on the fusion of multi-views of latent interests. Background Art

[0002] Recommendation systems are one of the most popular research topics in social networks due to their great potential in academic research and real life. For example, in e-commerce, recommendation systems can provide personalized information for specific users to reduce the selection cost when users face numerous products or services. In social media, recommendation systems can recommend friends with common hobbies or interests to users based on their personalized information. Traditional recommendation systems can generally be divided into methods based on collaborative filtering (CF), content-based methods, and hybrid methods, which mainly learn semantic features based on the interaction or behavior history records between users and items. In recent years, deep neural network models, such as convolutional neural networks, recurrent neural networks, and long short-term memory, have been adopted to learn the network structure and node representations, and have greatly improved the effects of user modeling and recommendation. A typical example is that the neural collaborative filtering model uses a multi-layer neural network to learn better user and item representations. On this basis, the model based on graph convolutional network (GCN) has become one of the most important basic models in deep learning-based recommendation systems due to its strong ability to learn network structure representations. This model can iteratively aggregate feature information from the local graph neighborhood with a non-Euclidean structure, and use high-order connectivity to alleviate the sparsity problem in recommendations. The representations of users and items will benefit from the additional information refined from the graph structure. Although these methods have achieved a certain degree of success, graph convolutional operations are essentially a special type of graph Laplacian smoothing, resulting in a serious over-smoothing problem in GCN models - after multiple graph convolutional operations, node representations gradually become similar, limiting the depth and size of the model. The LightGCN model proves that the performance of this model may be adversely affected by feature transformation and non-linear activation. This model only considers neighboring nodes, alleviates the over-smoothing problem to a certain extent by removing non-linearity and introducing a residual network structure, and removing the "self-loop" operation during the aggregation process. Further, in the graph convolutional operation, the IMP-GCN model also involves high-order adjacent users without common interests in the user embedding learning, and constructs a subgraph composed of users with similar interests and the items they interact with through an unsupervised subgraph generation module. Summary of the Invention

[0003] Aiming at the problem that the current recommendation system model lacks specific representation information, the present invention provides a recommendation method and system based on the fusion of multi-views of latent interests.

[0004] To achieve the above object, the present invention adopts the following technical solutions:

[0005] A recommendation system based on multi-view fusion of latent interests, including a multi-view generation module, a VS-LightGCN module, and a multi-view fusion module;

[0006] The multi-view generation module:

[0007] The original user-item interaction behavior matrix Among them, It is indicated that R is an m×n-dimensional matrix with each element being a real number; m and n respectively represent the number of users and items. If the element R corresponding to user u and item i in R ui is non-zero, it means that user u has interacted with item i, and there is an edge between user u and item i; otherwise, this R ui is zero;

[0008] Based on the original user-item interaction behavior matrix R, a bipartite graph G=(N, E) of users and items is used to provide structural information; where N represents the set of nodes, including the nodes of users u and items i, and E represents the set of edges;

[0009] Then, the singular value decomposition SVD is used to decompose the original user-item interaction behavior matrix R, expressed as: R≈P (t) Σ (t) (Q (t) ) T , where P (t) and Q (t) are the left and right singular matrices respectively, Σ (t) is a diagonal matrix composed of the first t largest singular values, and T is the vector transpose;

[0010] Each singular value is regarded as a latent interest topic in the recommendation system; user u and item i are respectively represented as singular vectors p (t) u ∈P (t) and q (t) i ∈Q (t) under the first t latent interest topics; under the first t latent interest topics, the similarity between p (t) u and q (t) i is regarded as the structural association score between user u and item i, expressed as:

[0011]

[0012] According to the structural correlation score score (t) ui of each user u and item i under the first t latent interest topics, a similarity matrix Among them, the corresponding elements of user u and item i

[0013] The original user-item interaction behavior matrix R is transformed into a different real-valued continuous matrix M determined by t latent interest topics (t) ; By selecting different values of t, different similarity matrices M are obtained (t) ;

[0014] In the original user-item interaction graph, when a user accidentally clicks on an item they are not interested in, this edge will introduce noise to the graph structure learning; Set a threshold η to remove the noisy edges with too low structure-related scores in the bipartite graph based on the similarity matrix M (t) The users and the items they interact with are grouped into different subgraphs to generate different views; For each view V (t) , its interaction matrix is defined as:

[0015]

[0016] where η is a threshold that controls the similarity of the edges in the process of generating two views;

[0017] When and only when the corresponding element of user u and item i in the similarity matrix M (t) is not less than the threshold η, and the corresponding element R in the original user-item interaction matrix R = 1, the corresponding element of user u and item i in view V ui is 1, otherwise it is 0; (t) Among them, the corresponding elements of user u and item i in view V

[0018] The VS-LightGCN module:

[0019] For each view generated by the multi-view generation module, the initial embeddings of user u and item i are respectively represented as e (0) u and e (0) i ; Learn the representations of user u and item i from different views of the latent interest topics; The views are independent of each other, and the nodes in a specific view subgraph can only propagate information with the neighbors within the subgraph; At the same time, through singular value decomposition SVD, user u or item i is represented as a t-dimensional embedding p (t) u and q (t) i under the first t latent interest topics, so that it contains the personalized features of user u or item i itself, and initialize the node embedding representation in view V (t) , that is, e (0) u = p (t) u, e (0) i = q (t) i ;

[0020] The propagation process is expressed as:

[0021]

[0022]

[0023] where and are static copies of e (0) u and e (0) i respectively, and do not participate in optimization and update; ensure that personalized features will not be lost during the propagation process, |N u | and |N i | represent the sets of neighbor nodes N u of user u and N i of item i respectively, k represents the number of iteration rounds;

[0024] Parallel training optimizes each view V (t) , and the loss function L of each view V (t) is expressed as:

[0025]

[0026] where δ is the sigmoid function, m represents the number of users, a is a neighbor item node of user u, b is a non-neighbor item node of user u, e u , e a , e b are the representation vectors of user u, item a, and item b respectively; T is the vector transpose, E (0) is the matrix composed of the initial representations of all user and item nodes, λ t is a hyperparameter in view V (t) , controlling the L2 regularization weight of E (0) ;

[0027] The multi-view fusion module: Dynamically adjusts the influence of each view on the recommendation result by the principal component control mechanism PCCM; In the user-item interaction graph, the similarity between user u and item i is denoted as s (0) ui ; For each generated view V (t) , after the information propagation and optimization of each view are completed, the similarity s of view V (t) is obtained through the formula (1)ui ,s (2) ui ,...,s (t) ui ; Through the principal component control mechanism PCCM, the fusion similarity between multiple views is defined as:

[0028]

[0029] where represents the fusion similarity between multiple views, the sigmoid() function maps the multi-view fusion score to [0, 1], and μ is a threshold used to control the view result fusion in the principal component control mechanism PCCM.

[0030] The principal component control mechanism PCCM is based on the fact that the latent interest of user u can only be accurately recommended when the user is willing to interact with a certain type of commodity.

[0031] A recommendation method based on multi-view fusion of latent interests includes the following steps:

[0032] Step 1, establish the original user-item interaction behavior matrix where represents that R is an m×n-dimensional matrix with each element being a real number; m and n are the numbers of users and items respectively. If the element R corresponding to user u and item i in R ui is non-zero, it means that user u has interacted with item i, and there is an edge between user u and item i; otherwise, this R ui is zero;

[0033] Step 2, based on the original user-item interaction behavior matrix R, construct a bipartite graph G=(N, E) of users and items to provide structural information; where N represents the set of nodes, including user and item nodes, and E represents the set of edges;

[0034] Step 3, use singular value decomposition SVD to decompose the original user-item interaction behavior matrix R, expressed as: R≈P (t) Σ (t) (Q (t) ) T , where P (t) and Q (t) are the left and right singular matrices respectively, Σ (t) is a diagonal matrix composed of the first t largest singular values, and T is the vector transpose;

[0035] Step 4, regard each singular value as a latent interest topic in the recommendation system; user u and item i are respectively represented by the first t latent interest topics as singular vectors p (t) u∈P (t) and q (t) i ∈Q (t) ; Under the first t potential interest topics, the similarity between p (t) u and q (t) i is regarded as the structural association score between user u and item i, denoted as:

[0036]

[0037] Step 5. According to the structural correlation scores score (t) ui of each user u and item i under the first t potential interest topics, generate a similarity matrix where the corresponding element of user u and item i

[0038] The original user-item interaction discrete matrix R is transformed into different real-valued continuous matrices M determined by t potential interest topics (t) ; By selecting different t values, different similarity matrices M can be obtained (t) ;

[0039] In the original user-item interaction graph, if a user accidentally clicks on an item that they are not interested in, this edge may introduce noise to the graph structure learning; by setting a threshold η to remove the noisy edges with too low structural correlation scores in the bipartite graph based on the similarity matrix M (t) , the users and the items they interact with are grouped into different subgraphs to generate different views; for each view V (t) , its interaction matrix is defined as:

[0040]

[0041] If and only if the corresponding element (t) of user u and item i in the similarity matrix M is not less than the threshold η, and the corresponding element R ui = 1 in the original user-item interaction matrix R, the corresponding element (t) of user u and item i in view V has a value of 1, otherwise 0;

[0042] The differences between the generated views can provide rich information for the model representation learning;

[0043] Step 6. For each view generated by the multi-view generation module, the traditional LightGCN model is improved, and the VS-LightGCN model is proposed to learn the graph structure embedding representations of users and items;

[0044] The initial embeddings of user u and item i are denoted as e (0) u and e (0) i ; learn the representations of users and items from different views of latent interest topics; to ensure that the model can learn the differential features of different views and avoid noise interference, each view is independent of each other, and the nodes in a specific view subgraph can only propagate information with the neighbors within the subgraph; at the same time, the user or item is represented as a t-dimensional embedding p (t) u and q (t) i under the first t latent interest topics through singular value decomposition (SVD), so that it contains personalized features about the characteristics of the user or item itself, and use them to initialize the node embedding representations in view V (t) , that is, e (0) u = p (t) u , e (0) i = q (t) i ;

[0045] The propagation process in VS-LightGCN is defined as follows:

[0046]

[0047]

[0048] where and are static copies of e (0) u and e (0) i respectively, and do not participate in the optimization update; this ensures that the personalized features will not be lost during the propagation process and effectively alleviates the problem of over-smoothing. |N u | and |N i | represent the sizes of the user set N u and the item set N i respectively, and k represents the number of iteration rounds;

[0049] Train and optimize each view in parallel. The loss function of each view V (t) is expressed as:

[0050]

[0051] where δ is the sigmoid function, m represents the number of users, a is a neighbor item node of user u, b is a non-neighbor item node of user u, eu , e a , e b are the representation vectors of user u, item a, and item b respectively. T is the vector transpose, and E (0) is the matrix composed of the initialized representations of all user and item nodes, and λ t is a hyperparameter in view V (t) to control the L2 regularization weight of E (0) ;

[0052] Step 7: Dynamically adjust the influence of each view on the recommendation result through the principal component control mechanism PCCM; the principal component control mechanism PCCM is based on the fact that the latent interest of user u can only be accurately recommended when the user is willing to interact with a certain type of item;

[0053] In the original user-item interaction graph R, the similarity between user u and item i is denoted as s (0) ui ; for each generated view V (t) , after the information propagation optimization of each view is completed, the similarity s of view V (t) is obtained through the formula (1) ui , s (2) ui ,..., s (t) ui ; through the principal component control mechanism PCCM, the fusion similarity between multiple views is defined as:

[0054]

[0055] where represents the fusion similarity between multiple views, the sigmoid() function maps the multi-view fusion score to [0,1], and μ is a threshold used to control the view result fusion in the principal component control mechanism PCCM.

[0056] Compared with the prior art, the present invention has the following advantages:

[0057] The DMV-GCN model proposed by the present invention is superior to all comparative baseline methods. In the three datasets of ml-100k, ml-1m, and lastfm, the present invention can achieve the highest improvement effect of 2.25% respectively compared with the sub-optimal model. The SVD of the user-item interaction matrix in the preprocessing process of the present invention is a one-time operation, which can effectively improve the reusability of the model and reduce the complexity of the model. In addition, the present invention continuously perceives personalized features by using the residual network structure and the static copy mechanism, so that they will not be completely covered during the model optimization process, which enables the model to effectively handle the over-smoothing problem. Description of the Drawings

[0058] Figure 1 It is a schematic diagram for multi-view generation based on potential interest topics;

[0059] Figure 2 It is a schematic diagram of the VS-LightGCN module. Detailed Implementation Manner

[0060] The present invention is implemented based on the PyTorch framework.

[0061] The detailed configuration of the model is shown in Table 1 below.

[0062] Table 1 Model Parameter Configuration Table

[0063]

[0064] For the proposed model, the vector dimension of each user and item node is d, the maximum propagation layer k value of the VS-LightGCN model is 3; two views are generated based on different potential interest topics, corresponding to the parameter t of the potential interest topic; the threshold η controls the similarity of the edges in the process of generating the two views, and the threshold μ controls the view result fusion in PCCM; the weight γ has three values in each dataset, corresponding to the three views (the original interaction view and the two generated views) respectively, and the weight λ t is the regularization term parameter of the VS-LightGCN model, and the model is optimized and learned using Adam. Each time, 2048 data are sampled for batch optimization, and the learning rate of the optimization function.

[0065] Example 1

[0066] A recommendation system based on multi-view fusion of potential interests includes a multi-view generation module, a VS-LightGCN module, and a multi-view fusion module;

[0067] The multi-view generation module:

[0068] The original user-item interaction behavior matrix Among them, it is represented that R is an m×n-dimensional matrix with each element being a real number; m and n respectively represent the number of users and items. If the element R corresponding to user u and item i in R ui is non-zero, it means that user u has interacted with item i, and there is an edge between user u and item i; otherwise, this R ui is zero;

[0069] Based on the original user-item interaction behavior matrix R, a bipartite graph G=(N,E) of user-item is used to provide structural information; where N represents the set of nodes, including nodes of user u and item i, and E represents the set of edges;

[0070] Then use singular value decomposition (SVD) to decompose the original user-item interaction behavior matrix R, which is expressed as: R≈P (t) Σ (t) (Q (t) ) T , where P (t) and Q (t) are the left and right singular matrices respectively, Σ (t) is a diagonal matrix composed of the first t largest singular values, and T is the vector transpose;

[0071] Regard each singular value as a potential interest topic in the recommendation system; user u and item i are respectively represented as singular vectors p (t) u ∈P (t) and q (t) i ∈Q (t) under the first t potential interest topics; under the first t potential interest topics, regard the similarity between p (t) u and q (t) i as the structural association score between user u and item i, which is expressed as:

[0072]

[0073] According to the structural correlation score score (t) ui of each user u and item i under the first t potential interest topics, generate a similarity matrix where the corresponding element of user u and item i

[0074] The original user-item interaction behavior matrix R is transformed into different real-valued continuous matrices M (t) determined by t potential interest topics; select different t values to obtain different similarity matrices M (t) ;

[0075] In the original user-item interaction graph, when a user accidentally clicks on an item they are not interested in, this kind of edge will bring noise to the graph structure learning; set a threshold η to remove the noise edges with too low structural correlation scores in the bipartite graph based on the similarity matrix M (t) , and the items interacted by users and users are grouped into different subgraphs to generate different views; for each view V (t), its interaction matrix is defined as:

[0076]

[0077] where η is a threshold that controls the similarity of edges in the process of generating two views;

[0078] When and only when the similarity matrix M (t) the corresponding element of user u and item i in is not less than the threshold η, and the corresponding element R in the original user-item interaction matrix R ui = 1, the corresponding element of user u and item i in view V (t) is 1, otherwise it is 0; Value is 1, otherwise it is 0;

[0079] The VS-LightGCN module:

[0080] For each view generated by the multi-view generation module, the initial embeddings of user u and item i are respectively represented as e (0) u and e (0) i ; learn the representations of user u and item i from different views of potential interest topics; the views are independent of each other, and the nodes in a specific view subgraph can only propagate information with neighbors within the subgraph; at the same time, represent user u or item i as a t-dimensional embedding p (t) u and q (t) i under the first t potential interest topics through singular value decomposition SVD, so that it contains personalized features about the characteristics of user u or item i itself, and initialize the node embedding representation in view V (t) , that is, e (0) u = p (t) u , e (0) i = q (t) i ;

[0081] The propagation process is expressed as:

[0082]

[0083]

[0084] where and are e (0) u and e (0) iStatic copies are not involved in optimization updates; this ensures that personalized features are not lost during dissemination,|N u |and|N i |represent the sets of neighbor nodes of user u and item i, respectively, denoted as N u and N i respectively, where k represents the number of iteration rounds;

[0085] Parallel training optimizes each view V (t) , and for each view V (t) , the loss function L is expressed as:

[0086]

[0087] where δ is the sigmoid function, m represents the number of users, a is a neighbor item node of user u, b is a non-neighbor item node of user u, and e u , e a , e b are the representation vectors of user u, item a, and item b respectively; T is the vector transpose, E (0) is the matrix composed of the initial representations of all user and item nodes, and λ t is a hyperparameter in view V (t) that controls the L2 regularization weight of E (0) ;

[0088] The multi-view fusion module: The principal component control mechanism PCCM dynamically adjusts the influence of each view on the recommendation result; in the user-item interaction graph, the similarity between user u and item i is denoted as s (0) ui ; for each generated view V (t) , after the information propagation and optimization of each view are completed, the similarity s of view V (t) is obtained through the formula (1) ui , s (2) ui ,..., s (t) ui ; through the principal component control mechanism PCCM, the fusion similarity between multiple views is defined as:

[0089]

[0090] where represents the fusion similarity between multiple views, the sigmoid() function maps the multi-view fusion score to [0,1], and μ is a threshold used to control the view result fusion in the principal component control mechanism PCCM.

[0091] The principal component control mechanism PCCM is based on the fact that accurate recommendations can only be made when the latent interests of user u are willing to interact with a certain type of commodity.

[0092] Example 2

[0093] A recommendation method based on multi-view fusion of latent interests

[0094] The user-item interaction behavior matrix R is in the form of:

[0095]

[0096] R ui ∈ {0, 1} represents the interaction relationship (purchase, click, browse, etc.) between user u and item i, and 0 and 1 correspond to no interaction and having had an interaction behavior between u and i respectively.

[0097] (1) Multi-view generation module

[0098] For the original interaction matrix R, the present invention decomposes the matrix based on the singular value decomposition (SVD) technique. The present invention generates 2 different views by using 64 latent interest topics and 52 latent interest topics respectively. Taking the use of 64 latent interest topics as an example, R is decomposed into:

[0099] R ≈ P (64) Σ (64) (Q (64) ) T

[0100] The present invention regards each singular value in Σ as a latent interest topic of the recommendation system. That is, user u and item i can be represented by these 64 latent interest topics, and are respectively expressed as singular vectors p (64) u ∈ P (64) and q (64) i ∈ Q (64) , and regards the similarity between p (64) u and q (64) i as the structural association score between user u and item i, defined as:

[0101]

[0102] According to score (64) ui , generate the similarity matrix where

[0103] By setting a threshold η to remove based on the similarity matrix M (64)For the noisy edges with too low structure-related scores in the bipartite graph, the users and the items they interact with are grouped into different subgraphs to generate different views. For view V (64) , its interaction matrix is defined as:

[0104]

[0105] Similarly, the present invention can obtain view V using 52 potential interest topics (52) .

[0106] (2) VS-LightGCN module

[0107] For the two generated potential interest topic views V (64) and V (52) , the singular vectors under each view are used to initialize the user and item node representations. Taking view V (64) as an example, e (0) u = p (64) u , e (0) i = q (64) i , and at the same time, static copies of e (0) u and e (0) i are generated and This static copy does not participate in the optimization update. The VS-LightGCN module is used to optimize and learn each node representation:

[0108]

[0109]

[0110] where |Nu| and |Ni| represent the sizes of the user set Nu and the item set Ni respectively, and k represents the number of iteration rounds.

[0111] Each view is optimized and trained in parallel. The loss function of view V (64) is expressed as:

[0112]

[0113] where δ is the sigmoid function, E (0) is the matrix composed of the initial representations of all user and item nodes, and λ 64 is a hyperparameter in view V (64) , controlling the L2 regularization weight of E (0) .

[0114] Similarly, the present invention can obtain the optimized representation of nodes under view V by using 52 potential interest topics. (52) The following is the node optimization representation under view V.

[0115] (3) Multi-view fusion module.

[0116] For the generated potential interest topic views V (64) and V (52) , after the information propagation optimization of each view is completed, the similarity s between user u and item i under views V and V (64) and V (52) is obtained through the formula (64) ui , s (52) ui . The similarity obtained by decomposing the original interaction matrix R is denoted as s (0) ui , and score fusion is performed through PCCM:

[0117]

[0118] where the sigmoid() function maps the multi-view fusion score to [0,1], and μ is a threshold.

[0119] It can be understood that the model automatically adjusts the model parameters and data representation according to the value of the loss function returned on all data, performs the next round of iteration, reduces the loss value of this round of iteration, and repeats the model iteration optimization process until the loss value tends to be stable or reaches the maximum number of iterations.

[0120] Among them, during the optimization process, the dataset is randomly divided into a training set and a test set according to a ratio of 8:1.

[0121] After the optimization process is completed, for any user u in the test set, by calculating the fusion similarity between each item and user u is obtained, and the top n items with the highest similarity scores are selected as the recommended item list for user u according to the similarity ranking from high to low.

[0122] It should be noted that some parameter settings and details during the model learning process. The detailed configuration of the model is shown in Table 1.

[0123] The vector dimension d of each user and item node of the proposed model is 64 dimensions, and the maximum propagation layer k value of the VS-LightGCN model is 3; two views V (64) and V (52); The threshold η controls the similarity of edges in the process of generating two views, and the threshold μ controls the view result fusion in PCCM; The weight γ has three values in each dataset, corresponding to three views (the original interaction view M (t) and the two generated views V (64) and V (52) , and the weight λ t is the regularization term parameter of the VS-LightGCN model. The model is optimized and learned using Adam. Each time, 2048 data samples are used for batch optimization, and the learning rate of the optimization function is adjusted.

[0124] The present invention is compared with four currently internationally advanced recommendation methods (NGCF, LightGCN, SGL, IMP-GCN). Experiments are carried out on three datasets: ml-100k, ml-1m, and lastfm. The commonly used Recall and NDCG in the field are used as evaluation indicators. The comparison results are shown in Table 2, where DMV-GVN is the method designed by the present invention. The underlined part represents the second-best result among the comparison methods, and Improv. represents the improvement effect of the method of the present invention compared with the second-best result.

[0125] Table 2 Comparison of test results

[0126]

[0127] As shown in Table 2, the DMV-GCN model proposed by the present invention is superior to all the comparison baseline methods. In the three datasets of ml-100k, ml-1m, and lastfm, the present invention can achieve the highest improvement effect of 2.25% respectively compared with the second-best model. The SVD of the user-item interaction matrix in the preprocessing process of the present invention is a one-time operation, which can effectively improve the reusability of the model and reduce the complexity of the model. In addition, the present invention continuously perceives personalized features by using the residual network structure and the static copy mechanism, so that they will not be completely covered in the model optimization process, which enables the model to effectively handle the over-smoothing problem.

[0128] The content not described in detail in the specification of the present invention belongs to the prior art well-known to those skilled in the art. Although the illustrative specific embodiments of the present invention are described above for the convenience of those skilled in the art to understand the present invention, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions created using the concept of the present invention are within the scope of protection.

Claims

1. A recommendation system based on multi-view fusion of potential interests, characterized in that: It includes a multi-view generation module, a VS-LightGCN module, and a multi-view fusion module; The multi-view generation module: Original user-item interaction behavior matrix Among them, It is indicated that R is an m×n-dimensional matrix with each element being a real number; m and n respectively represent the number of users and items. If the element R corresponding to user u and item i in R ui is non-zero, it means that user u has interacted with item i before, and there is an edge between user u and item i; otherwise, this R ui is zero; Based on the original user-item interaction behavior matrix R, a user-item bipartite graph G=(N, E) is used to provide structural information; where N represents the set of nodes, including nodes of users u and items i, and E represents the set of edges; Then, the Singular Value Decomposition (SVD) is used to decompose the original user-item interaction behavior matrix R, which is expressed as: R≈P (t) Σ (t) (Q (t) ) T , where P (t) and Q (t) are the left and right singular matrices respectively, Σ (t) is a diagonal matrix composed of the first t largest singular values, and T represents the vector transpose; Each singular value is regarded as a potential interest topic in the recommendation system; the user u and the item i are respectively represented as the singular vectors p (t) u ∈P (t) and q (t) i ∈Q (t) ; under the first t potential interest topics, the similarity between p (t) u and q (t) i is regarded as the structural association score between the user u and the item i, denoted as: According to the structure-related score score of each user u and item i under the top t potential interest topics (t) ui , a similarity matrix is generated where the corresponding element of user u and item i The original user-item interaction behavior matrix R is transformed into a different real-valued continuous matrix M determined by t latent interest topics (t) ; By selecting different values of t, different similarity matrices M are obtained (t) ; Set a threshold η to remove the noisy edges with too low structure-related scores in the bipartite graph based on the similarity matrix M (t) The users and the commodities they interact with are grouped into different subgraphs to generate different views; for each view V (t) , its interaction matrix is defined as: Among them, η is a threshold to control the similarity of edges in the process of generating two views; If and only if the similarity matrix M (t) The corresponding elements of user u and product i is not less than the threshold η, and the corresponding element R in the original user-product interaction matrix R ui = 1, view V (t) The corresponding elements of user u and product i The value is 1, otherwise it is 0; The VS-LightGCN module: For each view generated by the multi-view generation module, the initial embeddings of user u and item i are respectively denoted as e (0) u and e (0) i ; learn the representations of user u and item i from different views of latent interest topics; the views are independent of each other, and the nodes in a specific view subgraph can only propagate information with neighbors within the subgraph; at the same time, represent user u or item i as a t-dimensional embedding p (t) u and q (t) i under the first t latent interest topics through singular value decomposition (SVD), so that it contains personalized features about the characteristics of user u or item i itself, and initialize the node embedding representation in view V (t) , that is, e (0) u = p (t) u , e (0) i = q (t) i ; The propagation process is expressed as: where and are static copies of e (0) u and e (0) i respectively, and do not participate in optimization updates; ensure that personalized features are not lost during the propagation process, |N u | and |N i | represent the set of neighbor nodes N u of user u and the set of neighbor nodes N i of item i respectively, and k represents the number of iteration rounds; Parallel training optimizes each view V (t) For each view V (t) the loss function L is expressed as: Among them, δ is the sigmoid function, m represents the number of users, a is a neighbor commodity node of user u, b is a non-neighbor commodity node of user u, and e u 、e a 、e b are the representation vectors of user u, commodity a, and commodity b respectively; T is the vector transpose, and E (0) is the matrix composed of the initial representations of all user and commodity nodes, and λ t is a hyperparameter in view V (t) that controls the L2 regularization weight of E (0) . The multi-view fusion module: The principal component control mechanism (PCCM) is used to dynamically adjust the influence of each view on the recommendation result; in the user-item interaction graph, the similarity between user u and item i is denoted as s (0) ui ; For each generated view V (t) , after the information propagation optimization of each view is completed, through the formula the similarity s of view V (t) is obtained (1) ui , s (2) ui ,..., s (t) ui ; Through the principal component control mechanism PCCM, the fusion similarity between multiple views is defined as: Among them, represents the fusion similarity between multiple views. The sigmoid() function maps the multi-view fusion score to [0, 1], and μ is a threshold used to control the view result fusion in the principal component control mechanism PCCM.

2. The recommendation system based on potential interest multi-view fusion according to claim 1, characterized in that: The principal component control mechanism PCCM is based on the fact that the latent interest of user u can be accurately recommended only when the user is willing to interact with a certain type of item.

3. The recommendation method of a recommendation system based on multi-view fusion of potential interests according to claim 1, wherein: It includes the following steps: Step 1, establish the original user-item interaction behavior matrix wherein it is indicated that R is an m×n-dimensional matrix with each element being a real number; m and n respectively represent the number of users and items. If the element R ui corresponding to user u and item i in R is non-zero, it means that user u has interacted with item i before, and there is an edge between user u and item i; otherwise, this R ui is zero; Step 2, based on the original user-item interaction behavior matrix R, construct a user-item bipartite graph G=(N, E) to provide structural information; where N represents the set of nodes, including nodes of users u and items i, and E represents the set of edges; Step 3, use Singular Value Decomposition (SVD) to decompose the original user-item interaction behavior matrix R, expressed as: R≈P (t) Σ (t) (Q (t) ) T , where P (t) and Q (t) are the left and right singular matrices respectively, Σ (t) is a diagonal matrix composed of the first t largest singular values, and T represents vector transpose; Step 4: Consider each singular value as a potential interest topic in the recommendation system; the user u and the item i are respectively represented as singular vectors p (t) u ∈P (t) and q (t) i ∈Q (t) ; under the first t potential interest topics, consider the similarity between p (t) u and q (t) i as the structural association score between the user u and the item i, denoted as: Step 5. Generate a similarity matrix based on the structure correlation scores score of each user u and item i under the previous t potential interest topics (t) ui , where the original user-item interaction behavior matrix R is transformed into different real-valued continuous matrices M determined by t potential interest topics The corresponding element of user u and item i The original user-item interaction behavior matrix R is transformed into different real-valued continuous matrices M determined by t potential interest topics (t) ; Select different t values to obtain different similarity matrices M (t) ; In the original user-item interaction graph, when a user accidentally clicks on an item they are not interested in, this kind of edge will introduce noise to the graph structure learning; by setting a threshold η to remove the noisy edges with too low structure-related scores in the bipartite graph based on the similarity matrix M (t) The items interacted by users and users are grouped into different subgraphs to generate different views; for each view V (t) , its interaction matrix is defined as: Among them, η is a threshold to control the similarity of edges in the process of generating two views; If and only if the similarity matrix M (t) the corresponding element of user u and item i is not less than the threshold η, and the corresponding element R in the original user-item interaction matrix R ui = 1, then the corresponding element of user u and item i in view V (t) is 1, otherwise it is 0; ​ Step 6, for each view generated by the multi-view generation module, represent the initial embeddings of user u and item i as e (0) u and e (0) i ; learn the representations of user u and item i from different views of latent interest topics; the views are independent of each other, and nodes in a specific view subgraph can only propagate information with neighbors within the subgraph; at the same time, represent user u or item i as a t-dimensional embedding p (t) u and q (t) i under the first t latent interest topics through singular value decomposition (SVD), so that it contains personalized features about the characteristics of user u or item i itself, and initialize the node embedding representation in view V (t) , that is, e (0) u = p (t) u , e (0) i = q (t) i ; The propagation process in VS-LightGCN is expressed as: Among them and are static copies of e (0) u and e (0) i respectively, and do not participate in the optimization update; ensure that personalized features will not be lost during the propagation process, |N u | and |N i | respectively represent the set of neighbor nodes N u of user u and the set of neighbor nodes N i of commodity i, and k represents the number of iteration rounds; Parallel training optimizes each view V (t) For each view V (t) the loss function L is expressed as: Among them, δ is the sigmoid function, m represents the number of users, a is a neighbor commodity node of user u, b is a non-neighbor commodity node of user u, e u and e a and e b are the representation vectors of user u, commodity a, and commodity b respectively; T is the vector transpose, E (0) is the matrix composed of the initial representations of all user and commodity nodes, λ t is a hyperparameter in view V (t) that controls the L2 regularization weight of E (0) ; Step 7, dynamically adjust the influence of each view on the recommendation result through the principal component control mechanism PCCM; the principal component control mechanism PCCM is based on the fact that the latent interest of user u can be accurately recommended only when the user is willing to interact with a certain type of item; In the user-product interaction diagram, the similarity between user u and product i is denoted as s (0) ui ; For each generated view V (t) , after the optimization of information dissemination in each view is completed, through the formula the similarity s of view V (t) is obtained (1) ui , s (2) ui ,..., s (t) ui ; Through the principal component control mechanism PCCM, the fusion similarity between multiple views is defined as: Among them, represents the fusion similarity between multiple views. The sigmoid() function maps the multi-view fusion score to [0, 1], and μ is a threshold used to control the view result fusion in the principal component control mechanism PCCM.

Images