Group recommendation method and system based on multi-view hierarchical representation learning

Abstract

The invention discloses a group recommendation method and system based on multi-view hierarchical representation learning, and belongs to the technical field of deep learning. Most existing group recommendation methods are interactive-driven, that is, only user-item interaction records or group-item interaction records are used for deducing group preferences, scoring information is not sufficiently utilized, and only a single interaction view angle is considered in the modeling process, so that the group representation granularity is relatively coarse, and the interpretation of a recommendation result is not strong. On the basis, the group recommendation method based on multi-view hierarchical representation learning is provided, a user-item scoring matrix is divided into a high level, a middle level and a low level, collaborative modeling is carried out on users, groups and items from multiple views by comprehensively applying graph and hypergraph convolution, and through multi-view fusion, the user-item scoring matrix is classified into a high level, a middle level and a low level. According to the method, the comprehensive classification representation of the group and the comprehensive representation of the item are obtained, and then the recommendation level of the item is determined according to the matching condition of the item representation and the representation of each level of the group, so that a more explanatory high-quality recommendation result is provided for group members.

Description

Technical Field

[0001] The present invention belongs to the field of deep learning technology, and specifically relates to a group recommendation method and system based on multi-view hierarchical representation learning. Background Art

[0002] As an extension and extension of personalized recommendation, group recommendation usually infers group preferences by aggregating the interests of different members, and plays an important role in online shopping, social interaction, and life services. Group recommendation methods can be roughly divided into two stages: node representation modeling and member preference aggregation, aiming to generate a recommendation list for group members that meets the common preferences of the group. Graph neural networks, as a commonly used representation modeling method, are mainly used to capture local relationships in graph structures, while hypergraph networks can effectively model high-order associations and global collaborative relationships between nodes, and are often used to enhance the quality of node representation. The combined use of graph and hypergraph convolutions can model group preferences more comprehensively and accurately. In order to further improve the modeling quality, people introduce self-supervised learning into group recommendations to better capture the relationship between members and groups.

[0003] However, the current research on group recommendation methods is mainly based on historical interaction records, which does not fully utilize the rating information and cannot capture the fine-grained preferences of the group. Since only a single interaction perspective is considered, the analysis of the association between users, groups, and projects is not comprehensive enough, and other associations implicit in the information are ignored, making it difficult to obtain high-quality node representations. In addition, most of the existing group recommendation methods simply sort the projects according to the predicted scores, and the recommendation results lack interpretability, resulting in low user satisfaction. Summary of the invention

[0004] The purpose of the present invention is to solve the above problems and propose a group recommendation method and system based on multi-perspective hierarchical representation learning, which splits the user-item rating matrix into three levels: high, medium and low. It comprehensively uses graph and hypergraph convolution to collaboratively model users, groups and projects from multiple perspectives, and obtains a comprehensive hierarchical representation of the group and a comprehensive representation of the project through multi-perspective fusion. Then, the recommendation level of the project is determined according to the matching situation of the project representation and the representation of each level of the group, so as to provide group members with more interpretable and high-quality recommendation results.

[0005] In order to solve the above technical problems, the technical solution adopted by the present invention is:

[0006] A group recommendation method based on multi-view hierarchical representation learning includes the following steps:

[0007] Step 1: Split the user-item rating matrix into three levels: high, medium, and low. Based on the rating matrices of each level, a hypergraph network is constructed from the perspective of user-user association, and hypergraph convolution is used to learn user hierarchical representations.

[0008] Step 2: Construct a user-item bipartite graph based on the user-item interaction matrix, and use graph convolution to learn user and item representations from the perspective of user-item interaction;

[0009] Step 3: Combine the representations of each level under the user-user association perspective with the user representations under the user-item interaction perspective to form a comprehensive hierarchical representation of the user;

[0010] Step 4: Construct a group-project bipartite graph based on the group-project interaction matrix, and use graph convolution to learn group and project representations from the perspective of group-project interaction;

[0011] Step 5: Use the similarity measurement strategy to analyze and measure the consistency between the comprehensive hierarchical representation of users in the group and the group representation from the perspective of group-item interaction, and then determine the weights of the members in the group, and obtain the group hierarchical representation from the perspective of group-member association through weighted summation;

[0012] Step 6: Use the adaptive fusion mechanism to fuse the group hierarchical representation from the perspective of group-member association with the group representation from the perspective of group-item interaction to form a comprehensive hierarchical representation of the group;

[0013] Step 7: Based on the rating matrices of each level, a hypergraph network is constructed from the perspective of project-project association. The hierarchical representation of the project is learned using hypergraph convolution, and the overall representation of the project is obtained through weighted summation. The project representations from the three perspectives of project-project, user-project, and group-project are then integrated to form a comprehensive representation of the project.

[0014] Step 8: By calculating the inner product of the comprehensive representation of the group at each level and the comprehensive representation of the project, the predicted score of the project at each level is obtained, the maximum value is selected as the final predicted score of the project, and the recommended level of the project is determined according to the level corresponding to the final predicted score.

[0015] Furthermore, the implementation method of step 1 is as follows:

[0016] In order to obtain fine-grained user preferences, the user-item rating matrix is ​​split into three levels: high, medium, and low according to the size of the user's rating of the project. A hypergraph network is constructed to capture the association between users at each level. The hyperedge is composed of the current user and other users who have rated the same project with the user and have the same level. Then, the user embedding matrix U at the high, medium, and low levels is obtained by using hypergraph convolution. + , U o , U - , the specific calculation formula is as follows:

[0017]

[0018] in, Represent the degree matrix of nodes in each level of hypergraph, Respectively represent the affiliation between nodes and hyperedges at each level, Respectively represent the degree matrix of the hyperedge in each level of hypergraph, U∈R d represents the initial embedding matrix of the user, is a trainable parameter matrix;

[0019] The user u is obtained from the user embedding matrix at high, medium and low levels p Hierarchical characterization

[0020] Furthermore, the implementation method of step 2 is as follows:

[0021] In order to model the relationship between users and items, a bipartite graph from the user-item interaction perspective is constructed based on the user-item interaction matrix, and the user is initially embedded in the matrix U∈R m×d and the item initial embedding matrix I∈R n×d Stacked into the user-item initial embedding matrix And compare it with the adjacency matrix A u Input the graph convolution network together, and define the graph convolution of the (l+1)th layer as:

[0022]

[0023] in, represents the embedding matrix of the lth layer, represents the degree matrix of the node;

[0024] By averaging the embedding matrices obtained at each layer, we get the user-item embedding matrix from the user-item interaction perspective. The specific calculation formula is as follows:

[0025]

[0026] Among them, L is the number of convolutional layers, represents the embedding matrix of the lth layer;

[0027] Embedding matrix by user project Available user u p and Project i q Representation from the perspective of user-item interaction and

[0028] Furthermore, the implementation method of step 3 is as follows:

[0029] In order to characterize user preferences, the user hierarchical representation from the user-user association perspective is spliced ​​with the user representation from the user-item interaction perspective to form a comprehensive hierarchical representation of the user. The specific calculation formula is as follows:

[0030]

[0031] in, Represents user u p High, medium, and low level representations from the perspective of user-user relationships, Represents user u p Representation from the perspective of user-item interaction, is a trainable parameter matrix.

[0032] Furthermore, the implementation method of step 4 is as follows:

[0033] In order to model the relationship between groups and projects, a group-project bipartite graph is constructed based on the group-project interaction matrix, and the group is initially embedded in the matrix G∈R k×d and the item initial embedding matrix I∈R n×d Stacking into group item initial embedding matrix And compare it with the adjacency matrix A g They are input into the graph convolutional network together, and the graph convolutional network is used to learn the representation of groups and items from the perspective of group-item interaction. The graph convolution of the (l+1)th layer is defined as:

[0034]

[0035] in, represents the embedding matrix of the lth layer, represents the degree matrix of the node;

[0036] By averaging the embedding matrices obtained at each layer, we can obtain the group-item embedding matrix from the perspective of group-item interaction. The specific calculation formula is as follows:

[0037]

[0038] Among them, L is the number of convolutional layers, represents the embedding matrix of the lth layer;

[0039] Embedding matrix by group item Available group g t and Project i q Representation from the perspective of group-item interaction and

[0040] Furthermore, the implementation method of step 5 is as follows:

[0041] By aggregating the representations of group members at each level, we can obtain the hierarchical representation of the group from the perspective of group-member association. In order to make the weights of group members more reasonable, we use the similarity between the representations of different members at each level and the group representation. To determine the weight of members at each level The specific calculation formula is as follows:

[0042]

[0043]

[0044]

[0045]

[0046]

[0047]

[0048] in, Indicates group g t Representation from the perspective of group-item interaction, Respectively represent member u p The high, medium and low level representations of , Sim(·) is the cosine similarity function;

[0049] The high, medium and low level representations of the group members are weighted summed to obtain the representation of the group at the high, medium and low levels. The specific calculation formula is as follows:

[0050]

[0051]

[0052]

[0053] Furthermore, the implementation method of step 6 is as follows:

[0054] In order to obtain high-quality group representation, the group representation from the perspective of group-item interaction Respectively and their representation from the perspective of group-member association Adaptive fusion to obtain comprehensive representation of the group at three levels: high, medium, and low The specific calculation formula is as follows:

[0055]

[0056] in, and is the trainable parameter matrix, σ is the activation function, α t , and They represent the learnable weights respectively.

[0057] Furthermore, the implementation method of step 7 is as follows:

[0058] In order to obtain a fine-grained representation of project classification, the user-project rating matrix is ​​split into three levels: high, medium, and low according to the size of the user's rating of the project. A hypergraph network is constructed to capture the relationship between projects at each level. The hyperedge consists of the current project and other projects that have been rated by the user of the rated project and have the same level. Then, the user embedding matrix I at the three levels of high, medium, and low is obtained by using hypergraph convolution. + ,I o ,I - , the specific calculation formula is as follows:

[0059]

[0060] in, Represent the degree matrix of nodes in each level of hypergraph, Respectively represent the affiliation between nodes and hyperedges at each level, Represents the degree matrix of the hyperedge in each level of hypergraph, I∈R d represents the initial embedding matrix of the user, is a trainable parameter matrix;

[0061] From the project embedding matrix at high, medium and low levels, we can get project i from the perspective of project-project association: q Hierarchical characterization Then, the weight of each level of project representation is determined by calculating the proportion of each level of project score. And through weighted summation, we can get the overall representation of the project from the perspective of project-project association. The specific calculation formula is as follows:

[0062]

[0063]

[0064]

[0065]

[0066] in, Project i q The corresponding number of high scores, medium scores, and low scores;

[0067] In order to obtain information-rich project representation, the project representations from the perspectives of project-project association, user-project interaction, and group-project interaction are integrated to form a comprehensive project representation. Specifically, From the perspective of user-project interaction Perform the splicing operation, and then From the perspective of group-project interaction Adaptive fusion to obtain a comprehensive representation of the project q , the specific calculation formula is as follows:

[0068]

[0069] in, W +o- , W g and W f is the trainable parameter matrix and σ is the activation function.

[0070] Furthermore, the implementation method of step 8 is as follows:

[0071] By calculating the inner product between the comprehensive representation of the project and the representation of different levels of the group, the predicted score of the project at each level is obtained. The specific calculation formula is as follows:

[0072]

[0073] in, Represents the comprehensive representation of the group at high, medium and low levels, i q represents a comprehensive representation of the candidate items;

[0074] After obtaining the prediction scores at each level, Represents the maximum value of the prediction score, according to The corresponding relationship between the prediction scores of each level can determine the candidate item i q Recommended level The specific calculation formula is as follows:

[0075]

[0076] Use the total loss function Loss to optimize the method;

[0077] The total loss function consists of three parts: high-level loss function Loss + , Intermediate level loss function Loss o And low-level loss function Loss - ;

[0078] The calculation formula of each level loss function is as follows:

[0079]

[0080] Among them, σ is the activation function, G t Indicates group g t The sampled group is the training set of items, i.e. group g t With Project I q Interaction has occurred, but has not yet been established with project i q′ Interaction occurs;

[0081] The total loss function Loss is obtained by mixing the loss functions of each level:

[0082]

[0083] Among them, λ1 is the regularization coefficient and Θ represents the trainable parameters of the model.

[0084] A group recommendation system based on multi-view hierarchical representation learning includes the following modules:

[0085] User-user association perspective modeling module: split the user-item rating matrix into three levels: high, medium, and low according to the user's rating of the project, build a hypergraph network to capture the association relationship between users at each level, and obtain the user embedding matrix U at the three levels of high, medium, and low through hypergraph convolution. + , U o , U - , and then get user u p Hierarchical characterization

[0086] User-item interaction perspective modeling module: Based on the user-item interaction matrix, a bipartite graph from the user-item interaction perspective is constructed, and the user is initially embedded in the matrix U∈R m×d and the item initial embedding matrix I∈R n×d Stacked into the user-item initial embedding matrix And compare it with the adjacency matrix A u Input the graph convolutional network together to get the embedding matrix of each layer By averaging the embedding matrices obtained at each layer, we get the user-item embedding matrix from the user-item interaction perspective. Then get user u p and Project i q Representation from the perspective of user-item interaction and

[0087] User comprehensive hierarchical representation learning module: The user hierarchical representation from the user-user association perspective is spliced ​​with the user representation from the user-item interaction perspective to form a comprehensive hierarchical representation of the user.

[0088] Group-project interaction perspective modeling module: construct a group-project bipartite graph based on the group-project interaction matrix, and embed the group into the initial matrix G∈R k×d and the item initial embedding matrix I∈R n×d Stacking into group item initial embedding matrix And compare it with the adjacency matrix A g Input the graph convolutional network together to obtain the embedding matrix of each layer By averaging the embedding matrices obtained at each layer, we can obtain the group-item embedding matrix from the perspective of group-item interaction. Then we get group g t and Project i q Representation from the perspective of group-item interaction and

[0089] Group-member association perspective modeling module: Based on the similarity between the representations of different members at each level and the group representations under the group-item interaction perspective Determine the weight of members at each level The high, medium and low level representations of the group members are weighted summed to obtain the representation of the group at the high, medium and low levels.

[0090] Group comprehensive hierarchical representation learning module: Representation of groups from the perspective of group-item interaction Respectively and their representation from the perspective of group-member association Adaptive fusion to form a comprehensive representation of the group at three levels: high, medium, and low

[0091] Project comprehensive representation learning module: split the user-project rating matrix into three levels: high, medium, and low according to the user's rating of the project, build a hypergraph network to capture the correlation between projects at each level, and obtain the user embedding matrix I at the three levels of high, medium, and low through hypergraph convolution + ,I o ,I - , and then get the project i from the perspective of project-project association q Hierarchical characterization The weight of each level of project representation is determined by calculating the proportion of each level of project scores And through weighted summation, we can get the overall representation of the project from the perspective of project-project association. Will From the perspective of user-project interaction Perform the splicing operation, and then From the perspective of group-project interaction Adaptive fusion to form a comprehensive representation of the project q ;

[0092] Prediction and feedback module: By calculating the inner product between the comprehensive representation of the project and the representation of different levels of the group, the predicted score of the project at each level is obtained. make Represents the maximum value of the prediction score, according to The corresponding relationship between the prediction scores of each level can determine the candidate item i q Recommended level After obtaining the prediction score, the method is optimized using the total loss function obtained by mixing the loss functions of each level;

[0093] Recommendation list generation module: The system generates a corresponding project recommendation list and the recommendation level of each project in the list based on the comprehensive hierarchical representation of the group to be recommended.

[0094] Compared with the prior art, the present invention has the following beneficial effects:

[0095] 1. The present invention can model group preferences more accurately and meticulously, and by establishing a feedback mechanism with recommendation levels, provide group members with more valuable and explainable recommendation results, thereby enhancing the group recommendation effect and improving user satisfaction;

[0096] 2. The present invention can be used in various network applications such as e-commerce, travel, news information, financial services, etc. By improving the accuracy of pushed content, it helps groups efficiently find products or services that meet their needs, effectively improving user experience, commercial value and platform activity in various industries, thereby promoting the participation and growth rate of platform users. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] Figure 1 It is a schematic diagram of the process of the method recommended by the present invention;

[0098] Figure 2 It is a schematic diagram of the framework of the method recommended by the present invention;

[0099] Figure 3 Schematic diagram of the structure of the system recommended by the present invention. DETAILED DESCRIPTION

[0100] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0101] like Figure 1-3 As shown, a group recommendation method based on multi-view hierarchical representation learning includes the following steps:

[0102] Step 1: Split the user-item rating matrix into three levels: high, medium, and low. Use hypergraph convolution to obtain the user classification representation from the perspective of user-user association. The implementation method is as follows:

[0103] In order to obtain fine-grained user preferences, the user-item rating matrix is ​​split into three levels: high, medium, and low according to the size of the user's rating of the project. A hypergraph network is constructed to capture the association between users at each level. The hyperedge is composed of the current user and other users who have rated the same project with the user and have the same level. Then, the user embedding matrix U at the high, medium, and low levels is obtained by using hypergraph convolution. + , U o , U - , the specific calculation formula is as follows:

[0104]

[0105]

[0106]

[0107] in, Represent the degree matrix of nodes in each level of hypergraph, Respectively represent the affiliation between nodes and hyperedges at each level, Respectively represent the degree matrix of the hyperedge in each level of hypergraph, U∈R d represents the initial embedding matrix of the user, is a trainable parameter matrix;

[0108] The user u is obtained from the user embedding matrix at high, medium and low levels p Hierarchical characterization

[0109] Step 2: Construct a bipartite graph based on the user-item interaction matrix, and obtain the user and item representations from the user-item interaction perspective through graph convolution. The implementation method is as follows:

[0110] In order to model the relationship between users and items, a bipartite graph from the user-item interaction perspective is constructed based on the user-item interaction matrix, and the user is initially embedded in the matrix U∈R m×d and the item initial embedding matrix I∈R n×d Stacked into the user-item initial embedding matrix And compare it with the adjacency matrix A u Input the graph convolution network together, and define the graph convolution of the (l+1)th layer as:

[0111]

[0112] in, represents the embedding matrix of the lth layer, represents the degree matrix of the node;

[0113] By averaging the embedding matrices obtained at each layer, we get the user-item embedding matrix from the user-item interaction perspective. The specific calculation formula is as follows:

[0114]

[0115] Among them, L is the number of convolutional layers, represents the embedding matrix of the lth layer;

[0116] Embedding matrix by user project Available user u p and Project i q Representation from the perspective of user-item interaction and

[0117] Step 3: Combine the user hierarchical representation from the user-user association perspective with the user representation from the user-item interaction perspective to form a comprehensive hierarchical representation of the user. The implementation method is as follows:

[0118] In order to more comprehensively and accurately represent user preferences, the user hierarchical representation from the user-user association perspective is spliced ​​with the user representation from the user-item interaction perspective to form a comprehensive hierarchical representation of the user. The specific calculation formula is as follows:

[0119]

[0120]

[0121]

[0122] in, Represents user u p High, medium, and low level representations from the perspective of user-user relationships, Represents user u p Representation from the perspective of user-item interaction, is a trainable parameter matrix.

[0123] Step 4: Construct a bipartite graph based on the group-project interaction matrix, and obtain the group and project representations from the perspective of group-project interaction through graph convolution. The implementation method is as follows:

[0124] In order to model the relationship between groups and projects, a group-project bipartite graph is constructed based on the group-project interaction matrix, and the group is initially embedded in the matrix G∈R k×d and the item initial embedding matrix I∈Rn×d Stacking into group item initial embedding matrix And compare it with the adjacency matrix A g They are input into the graph convolutional network together, and the graph convolutional network is used to learn the representation of groups and items from the perspective of group-item interaction. The graph convolution of the (l+1)th layer is defined as:

[0125]

[0126] in, represents the embedding matrix of the lth layer, represents the degree matrix of the node;

[0127] By averaging the embedding matrices obtained at each layer, we can obtain the group-item embedding matrix from the perspective of group-item interaction. The specific calculation formula is as follows:

[0128]

[0129] Among them, L is the number of convolutional layers, represents the embedding matrix of the lth layer;

[0130] Embedding matrix by group item Available group g t and Project i q Representation from the perspective of group-item interaction and

[0131] Step 5: Perform weighted fusion on the representations of group members at each level to obtain the group representation from the perspective of group-member association. The implementation method is as follows:

[0132] By aggregating the representations of group members at each level, we can obtain the hierarchical representation of the group from the perspective of group-member association. In order to make the weights of group members more reasonable, we use the similarity between the representations of different members at each level and the group representation. To determine the weight of members at each level The specific calculation formula is as follows:

[0133]

[0134]

[0135]

[0136]

[0137]

[0138]

[0139] in, Indicates group g t Representation from the perspective of group-item interaction, Respectively represent member u p The high, medium and low level representations of , Sim(·) is the cosine similarity function;

[0140] The high, medium and low level representations of the group members are weighted summed to obtain the representation of the group at the high, medium and low levels. The specific calculation formula is as follows:

[0141]

[0142]

[0143]

[0144] Step 6: Adaptively fuse the group representation from the group-item interaction perspective with the group representation from the group-member association perspective to form a comprehensive representation of the group at each level. The implementation method is as follows:

[0145] In order to obtain high-quality group representation, the group representation from the perspective of group-item interaction Respectively and their representation from the perspective of group-member association Adaptive fusion to obtain comprehensive representation of the group at three levels: high, medium, and low The specific calculation formula is as follows:

[0146]

[0147]

[0148]

[0149] in, and is the trainable parameter matrix, σ is the activation function, α t , and They represent the learnable weights respectively.

[0150] Step 7: Obtain the overall representation of the project from the perspective of project-project association through hypergraph convolution and weighted fusion, and fuse it with the project representation from the perspective of user-project and group-project interaction to form a comprehensive representation of the project. The implementation method is as follows:

[0151] In order to obtain a fine-grained representation of project classification, the user-project rating matrix is ​​split into three levels: high, medium, and low according to the size of the user's rating of the project. A hypergraph network is constructed to capture the relationship between projects at each level. The hyperedge consists of the current project and other projects that have been rated by the user of the rated project and have the same level. Then, the user embedding matrix I at the three levels of high, medium, and low is obtained by using hypergraph convolution. + ,I o ,I - , the specific calculation formula is as follows:

[0152]

[0153] in, Represent the degree matrix of nodes in each level of hypergraph, Respectively represent the affiliation between nodes and hyperedges at each level, Represents the degree matrix of the hyperedge in each level of hypergraph, I∈R d represents the initial embedding matrix of the user, is a trainable parameter matrix;

[0154] From the project embedding matrix at high, medium and low levels, we can get project i from the perspective of project-project association: q Hierarchical characterization Then, the weight of each level of project representation is determined by calculating the proportion of each level of project score. And through weighted summation, we can get the overall representation of the project from the perspective of project-project association. The specific calculation formula is as follows:

[0155]

[0156]

[0157]

[0158]

[0159] in, Project i q The corresponding number of high scores, medium scores, and low scores;

[0160] In order to obtain information-rich project representation, the project representations from the perspectives of project-project association, user-project interaction, and group-project interaction are integrated to form a comprehensive project representation. Specifically, From the perspective of user-project interaction Perform the splicing operation, and then From the perspective of group-project interaction Adaptive fusion to obtain a comprehensive representation of the project q , the specific calculation formula is as follows:

[0161]

[0162]

[0163] in, W +o- , W g and W f is the trainable parameter matrix and σ is the activation function.

[0164] Step 8: The predicted scores of the projects at each level are obtained by calculating the inner product between the comprehensive representation of the project and the representations of different levels of the group, and the recommended level of the candidate project is determined according to the corresponding relationship. The implementation method is as follows:

[0165] By calculating the inner product between the comprehensive representation of the project and the representation of different levels of the group, the predicted score of the project at each level is obtained. The specific calculation formula is as follows:

[0166]

[0167]

[0168]

[0169] in, Represents the comprehensive representation of the group at high, medium and low levels, i q represents a comprehensive representation of the candidate items;

[0170] After obtaining the prediction scores at each level, Represents the maximum value of the prediction score, according to The corresponding relationship between the prediction scores of each level can determine the candidate item i q Recommended level The specific calculation formula is as follows:

[0171]

[0172] Use the total loss function Loss to optimize the method;

[0173] The total loss function consists of three parts: high-level loss function Loss + , Intermediate level loss function Loss o And low-level loss function Loss - ;

[0174] The calculation formula of each level loss function is as follows:

[0175]

[0176]

[0177]

[0178] Among them, σ is the activation function, G t Indicates group g t The sampled group is the training set of items, i.e. group g t With Project I q Interaction has occurred, but has not yet been established with project i q′ Interaction occurs;

[0179] The total loss function Loss is obtained by mixing the loss functions of each level:

[0180]

[0181] Among them, λ1 is the regularization coefficient and Θ represents the trainable parameters of the model.

[0182] A group recommendation system based on multi-view hierarchical representation learning includes the following modules:

[0183] User-user association perspective modeling module: split the user-item rating matrix into three levels: high, medium, and low according to the user's rating of the project, build a hypergraph network to capture the association relationship between users at each level, and obtain the user embedding matrix U at the three levels of high, medium, and low through hypergraph convolution. + , U o , U - , and then get user u p Hierarchical characterization

[0184] User-item interaction perspective modeling module: Based on the user-item interaction matrix, a bipartite graph from the user-item interaction perspective is constructed, and the user is initially embedded in the matrix U∈R m×d and the item initial embedding matrix I∈R n×d Stacked into the user-item initial embedding matrix And compare it with the adjacency matrix A u Input the graph convolutional network together to get the embedding matrix of each layer By averaging the embedding matrices obtained at each layer, we get the user-item embedding matrix from the user-item interaction perspective. Then get user u p and Project i q Representation from the perspective of user-item interaction and

[0185] User comprehensive hierarchical representation learning module: The user hierarchical representation from the user-user association perspective is spliced ​​with the user representation from the user-item interaction perspective to form a comprehensive hierarchical representation of the user.

[0186] Group-project interaction perspective modeling module: construct a group-project bipartite graph based on the group-project interaction matrix, and embed the group into the initial matrix G∈R k×d and the item initial embedding matrix I∈R n×d Stacking into group item initial embedding matrix And compare it with the adjacency matrix A g Input the graph convolutional network together to obtain the embedding matrix of each layer By averaging the embedding matrices obtained at each layer, we can obtain the group-item embedding matrix from the perspective of group-item interaction. Then we get group g t and Project i q Representation from the perspective of group-item interaction and

[0187] Group-member association perspective modeling module: Based on the similarity between the representations of different members at each level and the group representations under the group-item interaction perspective Determine the weight of members at each level The high, medium and low level representations of the group members are weighted summed to obtain the representation of the group at the high, medium and low levels.

[0188] Group comprehensive hierarchical representation learning module: Representation of groups from the perspective of group-item interaction Respectively and their representation from the perspective of group-member association Adaptive fusion to form a comprehensive representation of the group at three levels: high, medium, and low

[0189] Project comprehensive representation learning module: split the user-project rating matrix into three levels: high, medium, and low according to the user's rating of the project, build a hypergraph network to capture the correlation between projects at each level, and obtain the user embedding matrix I at the three levels of high, medium, and low through hypergraph convolution + ,I o ,I - , and then get the project i from the perspective of project-project association q Hierarchical characterization The weight of each level of project representation is determined by calculating the proportion of each level of project scores And through weighted summation, we can get the overall representation of the project from the perspective of project-project association. Will From the perspective of user-project interaction Perform the splicing operation, and then From the perspective of group-project interaction Adaptive fusion to form a comprehensive representation of the project q ;

[0190] Prediction and feedback module: By calculating the inner product between the comprehensive representation of the project and the representation of different levels of the group, the predicted score of the project at each level is obtained. make Represents the maximum value of the prediction score, according to The corresponding relationship between the prediction scores of each level can determine the candidate item i q Recommended level After obtaining the prediction score, the method is optimized using the total loss function obtained by mixing the loss functions of each level;

[0191] Recommendation list generation module: The system generates a corresponding project recommendation list and the recommendation level of each project in the list based on the comprehensive hierarchical representation of the group to be recommended.

Claims

1. A group recommendation method based on multi-view hierarchical representation learning, characterized in that: The following steps are involved: Step 1: Split the user-item rating matrix into three levels: high, medium, and low. Based on the rating matrices of each level, a hypergraph network is constructed from the perspective of user-user association, and hypergraph convolution is used to learn user hierarchical representations. Step 2: Construct a user-item bipartite graph based on the user-item interaction matrix, and use graph convolution to learn user and item representations from the perspective of user-item interaction; Step 3: Combine the representations of each level under the user-user association perspective with the user representations under the user-item interaction perspective to form a comprehensive hierarchical representation of the user; Step 4: Construct a group-project bipartite graph based on the group-project interaction matrix, and use graph convolution to learn group and project representations from the perspective of group-project interaction; Step 5: Use the similarity measurement strategy to analyze and measure the consistency between the comprehensive hierarchical representation of users in the group and the group representation from the perspective of group-item interaction, and then determine the weights of the members in the group, and obtain the group hierarchical representation from the perspective of group-member association through weighted summation; Step 6: Use the adaptive fusion mechanism to fuse the group hierarchical representation from the perspective of group-member association with the group representation from the perspective of group-item interaction to form a comprehensive hierarchical representation of the group; Step 7: Based on the rating matrices of each level, a hypergraph network is constructed from the perspective of project-project association. The hierarchical representation of the project is learned using hypergraph convolution, and the overall representation of the project is obtained through weighted summation. The project representations from the three perspectives of project-project, user-project, and group-project are then integrated to form a comprehensive representation of the project. Step 8: By calculating the inner product of the comprehensive representation of the group at each level and the comprehensive representation of the project, the predicted score of the project at each level is obtained, the maximum value is selected as the final predicted score of the project, and the recommended level of the project is determined according to the level corresponding to the final predicted score.

2. According to claim 1, a group recommendation method based on multi-view hierarchical representation learning is characterized in that: The implementation method of step 1 is as follows: In order to obtain fine-grained user preferences, the user-item rating matrix is ​​split into three levels: high, medium, and low according to the size of the user's rating of the project. A hypergraph network is constructed to capture the association between users at each level. The hyperedge is composed of the current user and other users who have rated the same project with the user and have the same level. Then, the user embedding matrix U at the high, medium, and low levels is obtained by using hypergraph convolution. + , U o , U - , the specific calculation formula is as follows: in, Represent the degree matrix of nodes in each level of hypergraph, Respectively represent the affiliation between nodes and hyperedges at each level, Respectively represent the degree matrix of the hyperedge in each level of hypergraph, U∈R d represents the initial embedding matrix of the user, is a trainable parameter matrix; The user u is obtained from the user embedding matrix at high, medium and low levels p Hierarchical characterization 3. The group recommendation method based on multi-view hierarchical representation learning according to claim 1, characterized in that: The implementation method of step 2 is as follows: In order to model the relationship between users and items, a bipartite graph from the user-item interaction perspective is constructed based on the user-item interaction matrix, and the user is initially embedded in the matrix U∈R m×d and the item initial embedding matrix I∈R n×d Stacked into the user-item initial embedding matrix And compare it with the adjacency matrix A u Input the graph convolution network together, and define the graph convolution of the (l+1)th layer as: in, represents the embedding matrix of the lth layer, represents the degree matrix of the node; By averaging the embedding matrices obtained at each layer, we get the user-item embedding matrix from the user-item interaction perspective. The specific calculation formula is as follows: Among them, L is the number of convolutional layers, represents the embedding matrix of the lth layer; Embedding matrix by user project Available user u p and Project i q Representation from the perspective of user-item interaction and 4. The group recommendation method based on multi-view hierarchical representation learning according to claim 1, characterized in that: The implementation method of step 3 is as follows: In order to characterize user preferences, the user hierarchical representation from the user-user association perspective is spliced ​​with the user representation from the user-item interaction perspective to form a comprehensive hierarchical representation of the user. The specific calculation formula is as follows: in, Represents user u p High, medium, and low level representations from the perspective of user-user relationships, Represents user u p Representation from the perspective of user-item interaction, is a trainable parameter matrix.

5. The group recommendation method based on multi-view hierarchical representation learning according to claim 1, characterized in that: The implementation method of step 4 is as follows: In order to model the relationship between groups and projects, a group-project bipartite graph is constructed based on the group-project interaction matrix, and the group is initially embedded in the matrix G∈R k×d and the item initial embedding matrix I∈R n×d Stacking into group item initial embedding matrix And compare it with the adjacency matrix A g They are input into the graph convolutional network together, and the graph convolutional network is used to learn the representation of groups and items from the perspective of group-item interaction. The graph convolution of the (l+1)th layer is defined as: in, represents the embedding matrix of the lth layer, represents the degree matrix of the node; By averaging the embedding matrices obtained at each layer, we can obtain the group-item embedding matrix from the perspective of group-item interaction. The specific calculation formula is as follows: Among them, L is the number of convolutional layers, represents the embedding matrix of the lth layer; Embedding matrix by group item Available group g t and Project i q Representation from the perspective of group-item interaction and 6. The group recommendation method based on multi-view hierarchical representation learning according to claim 1, characterized in that: The implementation method of step 5 is as follows: By aggregating the representations of group members at each level, we can obtain the hierarchical representation of the group from the perspective of group-member association. In order to make the weights of group members more reasonable, we use the similarity between the representations of different members at each level and the group representation. To determine the weight of members at each level The specific calculation formula is as follows: in, Indicates group g t Representation from the perspective of group-item interaction, Respectively represent member u p The high, medium and low level representations of , Sim(·) is the cosine similarity function; The high, medium and low level representations of the group members are weighted summed to obtain the representation of the group at the high, medium and low levels. The specific calculation formula is as follows:

7. The group recommendation method based on multi-view hierarchical representation learning according to claim 1, characterized in that: The implementation method of step 6 is as follows: In order to obtain high-quality group representation, the group representation from the perspective of group-item interaction Respectively and their representation from the perspective of group-member association Adaptive fusion to obtain comprehensive representation of the group at three levels: high, medium, and low The specific calculation formula is as follows: in, and is the trainable parameter matrix, σ is the activation function, α t , and They represent the learnable weights respectively.

8. The group recommendation method based on multi-view hierarchical representation learning according to claim 1, characterized in that: The implementation method of step 7 is as follows: In order to obtain a fine-grained representation of project classification, the user-project rating matrix is ​​split into three levels: high, medium, and low according to the size of the user's rating of the project. A hypergraph network is constructed to capture the relationship between projects at each level. The hyperedge consists of the current project and other projects that have been rated by the user of the rated project and have the same level. Then, the user embedding matrix I at the three levels of high, medium, and low is obtained by using hypergraph convolution. + ,I o ,I - , the specific calculation formula is as follows: in, Represent the degree matrix of nodes in each level of hypergraph, Respectively represent the affiliation between nodes and hyperedges at each level, Represents the degree matrix of the hyperedge in each level of hypergraph, I∈R d represents the initial embedding matrix of the user, is a trainable parameter matrix; From the project embedding matrix at high, medium and low levels, we can get project i from the perspective of project-project association: q Hierarchical characterization Then, the weight of each level of project representation is determined by calculating the proportion of each level of project score. And through weighted summation, we can get the overall representation of the project from the perspective of project-project association. The specific calculation formula is as follows: in, Project i q The corresponding number of high scores, medium scores, and low scores; In order to obtain information-rich project representation, the project representations from the perspectives of project-project association, user-project interaction, and group-project interaction are integrated to form a comprehensive project representation. Specifically, From the perspective of user-project interaction Perform the splicing operation, and then From the perspective of group-project interaction Adaptive fusion to obtain a comprehensive representation of the project q , the specific calculation formula is as follows: in, W +o- , W g and W f is the trainable parameter matrix and σ is the activation function.

9. The group recommendation method based on multi-view hierarchical representation learning according to claim 1, characterized in that: The implementation method of step 8 is as follows: By calculating the inner product between the comprehensive representation of the project and the representation of different levels of the group, the predicted score of the project at each level is obtained. The specific calculation formula is as follows: in, Respectively represent the comprehensive representation of the group at high, medium and low levels, i q represents a comprehensive representation of the candidate items; After obtaining the prediction scores at each level, Represents the maximum value of the prediction score, according to The corresponding relationship between the prediction scores of each level can determine the candidate item i q Recommended level The specific calculation formula is as follows: Use the total loss function Loss to optimize the method; The total loss function consists of three parts: high-level loss function Loss + , Intermediate level loss function Loss o And low-level loss function Loss - ; The calculation formula of each level loss function is as follows: Among them, σ is the activation function, G t Indicates group g t Sampling group - item training set, i.e. group g t With Project I q Interaction has occurred, but has not yet been established with project i q′ Interaction occurs; The total loss function Loss is obtained by mixing the loss functions of each level: Among them, λ1 is the regularization coefficient and Θ represents the trainable parameters of the model.

10. A group recommendation system based on multi-view hierarchical representation learning, characterized in that: Includes the following modules: User-user association perspective modeling module: split the user-item rating matrix into three levels: high, medium, and low according to the user's rating of the project, build a hypergraph network to capture the association relationship between users at each level, and obtain the user embedding matrix U at the three levels of high, medium, and low through hypergraph convolution. + , U o , U - , and then get user u p Hierarchical characterization User-item interaction perspective modeling module: Based on the user-item interaction matrix, a bipartite graph from the user-item interaction perspective is constructed, and the user is initially embedded in the matrix U∈R m×d and the item initial embedding matrix I∈R n×d Stacked into the user-item initial embedding matrix And compare it with the adjacency matrix A u Input the graph convolutional network together to get the embedding matrix of each layer By averaging the embedding matrices obtained at each layer, we get the user-item embedding matrix from the user-item interaction perspective. Then get user u p and Project i q Representation from the perspective of user-item interaction and User comprehensive hierarchical representation learning module: The user hierarchical representation from the user-user association perspective is spliced ​​with the user representation from the user-item interaction perspective to form a comprehensive hierarchical representation of the user. Group-project interaction perspective modeling module: construct a group-project bipartite graph based on the group-project interaction matrix, and embed the group into the initial matrix G∈R k×d and the item initial embedding matrix I∈R n×d Stacking into group item initial embedding matrix And compare it with the adjacency matrix A g Input the graph convolutional network together to obtain the embedding matrix of each layer By averaging the embedding matrices obtained at each layer, we can obtain the group-item embedding matrix from the perspective of group-item interaction. Then we get group g t and Project i q Representation from the perspective of group-item interaction and Group-member association perspective modeling module: Based on the similarity between the representations of different members at each level and the group representations under the group-item interaction perspective Determine the weight of members at each level The high, medium and low level representations of the group members are weighted summed to obtain the representation of the group at the high, medium and low levels. Group comprehensive hierarchical representation learning module: Representation of groups from the perspective of group-item interaction Respectively and their representation from the perspective of group-member association Adaptive fusion to form a comprehensive representation of the group at three levels: high, medium, and low Project comprehensive representation learning module: split the user-project rating matrix into three levels: high, medium, and low according to the user's rating of the project, build a hypergraph network to capture the correlation between projects at each level, and obtain the user embedding matrix I at the three levels of high, medium, and low through hypergraph convolution + ,I o ,I - , and then get the project i from the perspective of project-project association q Hierarchical characterization The weight of each level of project representation is determined by calculating the proportion of each level of project scores And through weighted summation, we can get the overall representation of the project from the perspective of project-project association. Will From the perspective of user-project interaction Perform the splicing operation, and then From the perspective of group-project interaction Adaptive fusion to form a comprehensive representation of the project q ; Prediction and feedback module: By calculating the inner product between the comprehensive representation of the project and the representation of different levels of the group, the predicted score of the project at each level is obtained. make Represents the maximum value of the prediction score, according to The corresponding relationship between the prediction scores of each level can determine the candidate item i q Recommended level After obtaining the prediction score, the method is optimized using the total loss function obtained by mixing the loss functions of each level; Recommendation list generation module: The system generates a corresponding project recommendation list and the recommendation level of each project in the list based on the comprehensive hierarchical representation of the group to be recommended.

Images