A recommendation method based on hypergraph motif optimization of multi-user representation

Abstract

The application provides a recommendation method for optimizing multi-user representation based on a hypergraph motif, and relates to the fields of knowledge graphs, recommendation systems and the like. The method constructs a user-item hypergraph based on the historical interactions of the user, maps it to a user adjacency graph, and uses it as the initial representation of the user implicit relationship network. Then, the hypergraph motif is extracted to learn the interaction mode of the multi-user, the implicit relationship network is reconstructed, and the item features in the hypergraph and the user association features in the implicit relationship network are fused to obtain a high-order representation of the user. Finally, the preference information of the user is propagated in the knowledge graph, and the representation of the item is aggregated. The method uses the hypergraph motif to learn the interaction mode of the multi-user, establishes a user implicit relationship network with direction and weight, can learn the interaction information of the multi-user, and then fuses the implicit relationship features of the user and the historical interaction item features to optimize the representation of the user and improve the performance of the recommendation system.

Description

Technical Field

[0001] This invention relates to a recommendation method based on hypergraph order optimization of multivariate user representations, belonging to the fields of knowledge graphs, recommendation systems, and related fields. Background Technology

[0002] Knowledge graphs are graph-structured knowledge bases. Due to their rich structured and semantic information, they alleviate the data sparsity and cold start problems in recommendation algorithms. Knowledge graphs are often used as external information to enrich the representation of items. At the same time, users have potential interests in each other, and learning the similarity between users is beneficial for modeling user representations. However, in many recommendation scenarios, it is not possible to obtain the explicit social relationships available between users. Therefore, it is necessary to use implicit relationships to establish connections between users.

[0003] Implicit relationships learn latent relationships between users by observing the similarity of behavioral interactions on third-party media to support recommendation systems. Existing methods for building implicit relationships only focus on the similarity between two users without considering the interaction patterns between multiple users. This results in implicit networks failing to include the interaction structure information of diverse users. For example, if one user has a clear guiding relationship with another user, then direction and weights can be established when learning the features of these two users, allowing the model to focus on more useful information.

[0004] Theoretically, simple structures in complex networks contain usable information, which corresponds to the interaction patterns of multiple users in implicit user relationship networks. However, existing methods usually build implicit relationship networks based on the similarity between pairs of users, making it difficult to establish information on direction and structure. Summary of the Invention

[0005] To address the shortcomings of existing recommendation methods in effectively modeling feature representations of diverse user interaction patterns, this invention provides a recommendation method based on hypergraph basalts to optimize diverse user representations, thereby improving the performance of recommendation systems. This invention establishes a mapping of implicit relationships between users by constructing a user-item interaction hypergraph, and learns interaction patterns among diverse users based on hypergraph basalts (a diverse structure within a hypergraph), establishing directions and weights in the implicit relationship network. To enrich user representations, a hierarchical convolutional network fusion of hypergraph convolutional networks and graph convolutional networks is designed to obtain features of implicit associations between items and users.

[0006] This invention is achieved through the following technical solution: a recommendation method based on hypergraph basalt optimization of multi-user representations. A user-item hypergraph is constructed based on the user's historical interactions, and this hypergraph is mapped to a user adjacency graph, serving as the initial representation of the user's implicit relationship network. Then, hypergraph basalts are extracted to learn the interaction patterns of multi-users, reconstructing the implicit relationship network. Item features from the hypergraph and user association features from the implicit relationship network are then fused to obtain a higher-order representation of the user. Finally, user preference information is propagated in a knowledge graph, item representations are aggregated, and the interaction probability between users and items is predicted based on the feature vector representations of users and items to complete the recommendation task.

[0007] Preferably, the specific steps include the following:

[0008] Step 1: Given a knowledge graph G = (E, R), where G represents the knowledge graph, E represents the entity set in the knowledge graph, and R represents the relation set, its triple form is G = {e, r, e}. ′ |e,e ′ ∈E,r∈R}, where e represents the head entity in the triple, e ′ Representing the tail entity in the triple, r represents the relationship between entities, constructing the user-project interaction hypergraph HG:;

[0009] Step 2: Extract the hypergraph motif hm from HG and reconstruct G. u The structure;

[0010] Step 3: Fusing the hypergraph HG and the implicit relational network G using a hierarchical graph convolutional network. u The feature information is used to obtain a higher-order representation of the user;

[0011] Step 4: Disseminate user preference information on the knowledge graph;

[0012] Step 5: Predict the probability of user interaction with the project.

[0013] Preferably, Step 1 further includes the following steps: Step 1.1, for user u, establish a hyperedge connecting it to all its historical interaction items:

[0014]

[0015] Where R u This represents the hyperedge of user u, where → indicates a connection relationship. This indicates the user's historical interaction items;

[0016] Step 1.2: Iterate through all users using the method in Step 1.1 to complete the construction of HG;

[0017] Step 1.3: Map the user-item interaction hypergraph HG to the user adjacency graph G. u :

[0018] Step 1.3.1: In the user-project interaction hypergraph, each hyperedge represents a user. The following criteria are used to determine whether the hyperedges in the user-project hypergraph are connected:

[0019]

[0020] Here, I(·) is a decision function, which is 1 if there are items between the hyperedges, and 0 otherwise. and They represent user u respectively i and user u j The superedges are denoted by ∩, which represents the intersection.

[0021] Step 1.3.2: For For users, the following function is used to calculate the similarity between users:

[0022]

[0023] Among them, v k sim(·) represents a single interactive item in the intersection, and sim(·) represents the similarity score of a single common interactive item. The similarity between users is calculated by score(·) and used as the weight of the edge.

[0024] Step 1.4: Connect the user adjacency graph G u This serves as the initialization of the user's implicit relationship network.

[0025] Preferably, Step 2 further includes the following steps: Step 2, extracting the hypergraph motif hm from HG and reconstructing G. u The structure;

[0026] Step 2.1: Calculate the reliability of the hypergraph order based on its interaction patterns and number of interactions.

[0027]

[0028] Where Re represents the reliability score of the hypergraph motif, |R u | represents the total number of nodes connected by the hyperedge, hm represents the hypergraph radix, and I hm This indicates whether there are common interactive items in the hypergraph's basic order; 1 indicates if they exist, and 0 otherwise.

[0029] Step 2.1.1: Set a threshold and filter unreliable hypergraph motifs based on reliability scores;

[0030] Step 2.2: Calculate similarity weights and establish directions based on multi-user interactions;

[0031] Step 2.2.1: Since the two users interact with more than one hypergraph motif, normalization is performed based on all hypergraph motifs:

[0032]

[0033] Where score(·) is the similarity score based on binary users in Step 1.3.2, and Re is the hypergraph order reliability score. It is a normalization function based on the attention mechanism, used to balance the values ​​between binary and ternary relations, s ij It is the end user u i and user u j Similarity score between them;

[0034] Step 2.2.2: Select and establish directionality based on the weight of the relevant hypergraph motif types;

[0035] Step 2.3, Traverse G u Find all edges in HG and locate all related hypergraph base orders hm;

[0036] Step 2.4: Reconstruct G according to Step 2.2 u The structure.

[0037] Preferably, Step 3 further includes the following steps: Step 3.1, using HG as input to the hypergraph convolutional network and utilizing the correlation between items to obtain the embedding of the user representation;

[0038] Step 3.1.1, from the hyperedge (R) u The connected nodes (user history interaction item set I) u To perform further propagation, its convolution operation is defined as:

[0039]

[0040] Where σ(·) is the activation function, P represents the trainable weight matrix between the (l)th layer and the (l+1)th layer, ∈ represents the relevant hyperedge, and W represents the bias matrix;

[0041] Step 3.1.2: Multiple hypergraph convolutional layers may lead to numerical instability, therefore symmetric normalization is added, and it is represented in matrix form as follows:

[0042]

[0043] Where σ(·) is the activation function, D and B are the degree matrices of nodes and hyperedges in the hypergraph, and H is the degree of the hypergraph nodes and hyperedges.T This indicates the transpose of H;

[0044] Step 3.1.3: Obtain the user embedding representation by aggregating the embeddings of items on the hyperedge:

[0045]

[0046] Where u represents the user's feature vector representation, σ(·) is the activation function, and f agg (·) denotes an aggregate function, i u It is the representation of the user's historical interaction items in the hypergraph;

[0047] Step 3.2, Obtain user information in G u Embedded representation in;

[0048] Step 3.2.1: Obtain the user's embedding representation based on the hypergraph convolutional network and use it as the input to the graph convolutional network. The convolution operation is defined as follows:

[0049]

[0050] Where N(i) represents the neighbors of node i, c ji It is the product of the square roots of the degree of the nodes, e ji σ is the scalar weight from node j to node i, and σ(·) is an activation function;

[0051] Step 3.3: Use the user representation output in the graph convolutional network as the input in the hypergraph HG, and repeat steps 3.1 and 3.2 until the required number of hierarchical convolutional layers is met;

[0052] Step 3.4: Obtain the user's final embedded representation u.

[0053] Preferably, Step 4 further includes the following steps:

[0054] Step 4.1: To better understand users' knowledge preferences, a function is used to calculate a score between users and relationships, thus determining the importance of different relationships in the knowledge graph to users.

[0055]

[0056] Where g(·) represents the calculation function for user preference for relation type, I u Let N(i) represent the set of a user's historical interaction items, and let r represent the set of neighbors of i in the knowledge graph. i This indicates the relation type when associating with a neighbor, exp(u·r i ) Calculate user u and r iThe inner product, where u and r represent the feature vector representations of the user and the relation, respectively. It is the user's final preference weight for the relationship;

[0057] Step 4.2: Calculate the user's preference weight for the project:

[0058]

[0059] Among them, |I u | represents the number of user interaction items, |v i | represents v in HG i The node degree, h(·), is used to calculate the preference weight of user u for item i.

[0060] Step 4.3: Aggregate item representations in the knowledge graph based on the attention mechanism;

[0061] Step 4.3.1: Determine the ratio of user preference weights for relationships based on the user's preference weights for items. For items with high interaction rates, aggregate neighboring entities more evenly.

[0062] Step 4.3.2: Sample entity features from the project's local domain and aggregate the project's multi-hop neighbors to explore more information in the knowledge graph. The project aggregation is represented as follows:

[0063]

[0064] Where Relu(·) represents the activation function, and N(i) represents item i. u The neighbor set is given by n, where n represents the number of sampled neighbors, W and b are the bias matrix and bias term, respectively, and v represents the eigenvector representation of the item.

[0065] Preferably, Step 5 further includes the following steps:

[0066] Step 5.1: Calculate the interaction probability between the project's representation v and the user's representation u, as shown below:

[0067]

[0068] in, This represents the predictive scoring function;

[0069] Step 5.2: Iterate through all possible user item pairs using a negative sampling strategy. The loss function is as follows:

[0070]

[0071] in, It is the L2-regularized loss function, λ, λ uand λ i It is a balancing hyperparameter. It is cross-entropy loss. It is the user aggregation loss. It is the aggregate loss of the project. This represents the user-level hierarchical convolutional aggregation of layer l. This represents the project convolutional aggregation propagation at layer h, y uv and These are the actual interaction and the predicted interaction probability, respectively.

[0072] Step 5.3, Model Performance Evaluation:

[0073] Step 5.3.1 Evaluation Metrics: For click-through rate prediction scenarios, accuracy (AUC), ACC, and F1 scores are used to evaluate model performance; for Top-K recommendation scenarios, Precision@K and Recall@K are used to evaluate model performance.

[0074] This invention first addresses the problems of optimizing user feature representations and learning multi-user interaction information by designing a recommendation method based on hypergraph emphases to optimize multi-user representations. This method constructs a user-item interaction hypergraph and maps it to a user implicit relationship network. Then, it extracts hypergraph emphases to reconstruct the structure of the implicit relationship network, establishing directions and weights. Feature information learned from the hypergraph convolutional network and the graph convolutional network is fused using a hierarchical convolutional approach to enrich the user's feature representation. Finally, user preferences are propagated on a knowledge graph to obtain item representations. Based on user and item embeddings, the interaction probability between users and items is predicted for recommendation purposes.

[0075] This invention provides a recommendation method based on hypergraph radix optimization of multivariate user representations. It has the following beneficial effects:

[0076] Existing recommendation methods often establish implicit associations between users based on the similarity of behaviors between binary users, neglecting the interaction pattern information of multiple users. The method proposed in this invention assists in the feature representation of users by learning the structural information of implicit interactions between multiple users. This method establishes a mapping between a user-item hypergraph and a user implicit relationship network, learns the interaction patterns of multiple users through hypergraph primacy, and designs a hierarchical convolution method to fuse item features and implicit relationship features between users. Compared with existing recommendation methods, this method optimizes the feature representation of users through implicit interaction relationships between multiple users to improve the performance of the recommendation system. It can also improve the interpretability of recommendation results from the attribute association of items and the implicit association between users. Experimental comparisons with other state-of-the-art methods demonstrate that this method can achieve excellent performance in different recommendation scenarios. Attached Figure Description

[0077] Figure 1 This is a flowchart of the recommended method proposed in this invention. Detailed Implementation

[0078] The present invention will be further described below with reference to the embodiments.

[0079] Example 1;

[0080] like Figure 1 As shown, this invention provides a recommendation method based on hypergraph basalt optimization of multi-user representations. The method includes constructing a user-item hypergraph from users' historical interactions, mapping it to a user adjacency graph as the initial representation of the user's implicit relationship network. Then, hypergraph basalts are extracted to learn the interaction patterns of multi-users, reconstructing the implicit relationship network, and fusing item features from the hypergraph with user association features from the implicit relationship network to obtain a higher-order representation of the user. Finally, user preference information is propagated in a knowledge graph, item representations are aggregated, and the interaction probability between users and items is predicted based on the feature vector representations of users and items to complete the recommendation task. By optimizing the user's feature representation through implicit interaction relationships among multi-users, the performance of the recommendation system is improved. Furthermore, the interpretability of the recommendation results can be enhanced by considering the attribute associations of items and the implicit associations between users. Experimental comparisons with other state-of-the-art methods demonstrate that this method achieves excellent performance in various recommendation scenarios.

[0081] The specific steps are as follows:

[0082] Step 1: Given a knowledge graph G = (E, R), where G represents the knowledge graph, E represents the entity set in the knowledge graph, and R represents the relation set, its triple form is G = {e, r, e}. ′ |e,e ′ ∈E,r∈R}, where e represents the head entity in the triple, e ′ Let r represent the tail entity in the triple, and r represent the relationship between entities. Construct the user-project interaction hypergraph HG:

[0083] Step 2: Extract the hypergraph motif hm from HG and reconstruct G. u The structure;

[0084] Step 3: Fusing the hypergraph HG and the implicit relational network G using a hierarchical graph convolutional network. u The feature information is used to obtain a higher-order representation of the user;

[0085] Step 4: Disseminate user preference information on the knowledge graph;

[0086] Step 5: Predict the probability of user interaction with the project.

[0087] Step 1 also includes the following steps:

[0088] Step 1.1: For user u, create a hyperedge that connects to all its historical interaction items:

[0089]

[0090] Where R u This represents the hyperedge of user u, where → indicates a connection relationship. This indicates the user's historical interaction items;

[0091] Step 1.2: Iterate through all users using the method in Step 1.1 to complete the construction of HG;

[0092] Step 1.3: Map the user-item interaction hypergraph HG to the user adjacency graph G. u :

[0093] Step 1.3.1: In the user-project interaction hypergraph, each hyperedge represents a user. The following criteria are used to determine whether the hyperedges in the user-project hypergraph are connected:

[0094]

[0095] Here, I(·) is a decision function, which is 1 if there are items between the hyperedges, and 0 otherwise. and They represent user u respectively i and user u j The superedges are denoted by ∩, which represents the intersection.

[0096] Step 1.3.2: For For users, the following function is used to calculate the similarity between users:

[0097]

[0098] Among them, v k sim(·) represents a single interactive item in the intersection, and sim(·) represents the similarity score of a single common interactive item. The similarity between users is calculated by score(·) and used as the weight of the edge.

[0099] Step 1.4: Connect the user adjacency graph G u This serves as the initialization of the user's implicit relationship network.

[0100] Step 2: Extract the hypergraph motif hm from HG and reconstruct G. u The structure;

[0101] Step 2.1: Calculate the reliability of the hypergraph order based on its interaction patterns and number of interactions.

[0102]

[0103] Where Re represents the reliability score of the hypergraph motif, |R u | represents the total number of nodes connected by the hyperedge, hm represents the hypergraph radix, and I hm This indicates whether there are common interactive items in the hypergraph's basic order; 1 indicates if they exist, and 0 otherwise.

[0104] Step 2.1.1: Set a threshold and filter unreliable hypergraph motifs based on reliability scores;

[0105] Step 2.2: Calculate similarity weights and establish directions based on multi-user interactions;

[0106] Step 2.2.1: Since the two users interact with more than one hypergraph motif, normalization is performed based on all hypergraph motifs:

[0107]

[0108] Where score(·) is the similarity score based on binary users in Step 1.3.2, and Re is the hypergraph order reliability score. It is a normalization function based on the attention mechanism, used to balance the values ​​between binary and ternary relations, s ij It is the end user u i and user u j Similarity score between them;

[0109] Step 2.2.2: Select and establish directionality based on the weight of the relevant hypergraph motif types;

[0110] Step 2.3, Traverse G u Find all edges in HG and locate all related hypergraph base orders hm;

[0111] Step 2.4: Reconstruct G according to Step 2.2 u The structure.

[0112] Step 3 also includes the following steps:

[0113] Step 3.1: Use HG as input to the hypergraph convolutional network and utilize the correlation between items to obtain the embedding of the user representation;

[0114] Step 3.1.1, from the hyperedge (R) u The connected nodes (user history interaction item set I) u To perform further propagation, its convolution operation is defined as:

[0115]

[0116] Where σ(·) is the activation function, P represents the trainable weight matrix between the (l)th layer and the (l+1)th layer, ∈ represents the relevant hyperedge, and W represents the bias matrix;

[0117] Step 3.1.2: Multiple hypergraph convolutional layers may lead to numerical instability, therefore symmetric normalization is added, and it is represented in matrix form as follows:

[0118]

[0119] Where σ(·) is the activation function, D and B are the degree matrices of nodes and hyperedges in the hypergraph, and H is the degree of the hypergraph nodes and hyperedges. T This indicates the transpose of H;

[0120] Step 3.1.3: Obtain the user embedding representation by aggregating the embeddings of items on the hyperedge:

[0121] u=σ(f agg (i u (8)

[0122] Where u represents the user's feature vector representation, σ(·) is the activation function, and f agg (·) denotes an aggregate function, i u It is the representation of the user's historical interaction items in the hypergraph;

[0123] Step 3.2, Obtain user information in G u Embedded representation in;

[0124] Step 3.2.1: Obtain the user's embedding representation based on the hypergraph convolutional network and use it as the input to the graph convolutional network. The convolution operation is defined as follows:

[0125]

[0126] Where N(i) represents the neighbors of node i, c ji It is the product of the square roots of the degree of the nodes, e ji σ is the scalar weight from node j to node i, and σ(·) is an activation function;

[0127] Step 3.3: Use the user representation output in the graph convolutional network as the input in the hypergraph HG, and repeat steps 3.1 and 3.2 until the required number of hierarchical convolutional layers is met;

[0128] Step 3.4: Obtain the user's final embedded representation u.

[0129] Step 4 also includes the following steps:

[0130] Step 4: Disseminate user preference information on the knowledge graph;

[0131] Step 4.1: To better understand users' knowledge preferences, a function is used to calculate a score between users and relationships, thus determining the importance of different relationships in the knowledge graph to users.

[0132]

[0133] Where g(·) represents the calculation function for user preference for relation type, I u Let N(i) represent the set of a user's historical interaction items, and let r represent the set of neighbors of i in the knowledge graph. i This indicates the relation type when associating with a neighbor, exp(u·r i ) Calculate user u and r i The inner product, where u and r represent the feature vector representations of the user and the relation, respectively. It is the user's final preference weight for the relationship;

[0134] Step 4.2: Calculate the user's preference weight for the project:

[0135]

[0136] Among them, |I u | represents the number of user interaction items, |v i | represents v in HG i The node degree, h(·), is used to calculate the preference weight of user u for item i.

[0137] Step 4.3: Aggregate item representations in the knowledge graph based on the attention mechanism;

[0138] Step 4.3.1: Determine the ratio of user preference weights for relationships based on the user's preference weight for each item. For items with high interaction rates, aggregate neighboring entities more evenly. For example, most people like highly rated movies, but niche movies may reflect a wider range of user interests.

[0139] Step 4.3.2: Sample entity features from the project's local domain and aggregate the project's multi-hop neighbors to explore more information in the knowledge graph. The project aggregation is represented as follows:

[0140]

[0141] Where Relu(·) represents the activation function, and N(i) represents item i. u The neighbor set is given by n, where n represents the number of sampled neighbors, W and b are the bias matrix and bias term, respectively, and v represents the eigenvector representation of the item.

[0142] Step 5 also includes the following steps:

[0143] Step 5: Predict the probability of user interaction with the project;

[0144] Step 5.1: Calculate the interaction probability between the project's representation v and the user's representation u, as shown below:

[0145]

[0146] in, This represents the predictive scoring function;

[0147] Step 5.2: Iterate through all possible user item pairs using a negative sampling strategy. The loss function is as follows:

[0148]

[0149] in, L2-regularization loss function, λ, λ u and λ i It is a balancing hyperparameter. It is cross-entropy loss. It is the user aggregation loss. It is the aggregate loss of the project. This represents the user-level hierarchical convolutional aggregation of layer l. This represents the project convolutional aggregation propagation at layer h, y uv and These are the actual interaction and the predicted interaction probability, respectively.

[0150] Step 5.3, Model Performance Evaluation:

[0151] Step 5.3.1 Evaluation Metrics: For click-through rate prediction scenarios, accuracy AUC (Area Under Curve), ACC (Accuracy), and F1 score are used to evaluate model performance; for Top-K recommendation scenarios, Precision@K and Recall@K are used to evaluate model performance.

[0152] Step 5.3.2, Experimental Datasets: Experiments were conducted using commonly used public datasets in the recommender system field, such as Movielens-1M and Last.FM; detailed information is shown in Table 1.

[0153] Table 1. Detailed information about the dataset.

[0154]

[0155] Step 5.3.3, Experimental Results: The training, evaluation, and test set ratio for each dataset was 8:1:1. Each experiment was repeated 10 times, and the average performance was reported. For the hit rate prediction scenario, the experimental results for accuracy (AUC), accuracy (ACC), and F1 score are shown in Table 2. For the Top-K recommendation scenario, the experimental results for Precision@K and Recall@K are shown in Table 3.

[0156] Table 2. CTR Click Prediction Results

[0157]

[0158] Table 3. Top-K Recommendation Results

[0159]

[0160]

[0161] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A recommendation method based on hypergraph basalt optimization of multivariate user representations, characterized in that: A user-item hypergraph is constructed based on the user's historical interactions, and it is mapped to a user adjacency graph as the initial representation of the user's implicit relationship network. Then, the hypergraph tuples are extracted to learn the interaction patterns of multiple users, the implicit relationship network is reconstructed, and the item features in the hypergraph and the user association features in the implicit relationship network are fused to obtain the user's high-order representation. Finally, the user's preference information is propagated in the knowledge graph, the item representation is aggregated, and the interaction probability between the user and the item is predicted based on the feature vector representations of the user and the item to complete the recommendation task. The specific steps are as follows: Step 1, Given the knowledge graph as follows ,in Representing a knowledge graph, Represents the entity set in a knowledge graph. Representing a relation set, its triple form is as follows: ,in, This represents the head entity in the triple. Represents the tail entity in a triple. Represent the relationships between entities and construct a user-project interaction hypergraph. : Step 1 also includes the following steps: Step 1.1, for users Establish a hyperedge that connects to all its historical interaction items: in User The hyperedge representation, Indicates a connection relationship. This indicates the user's historical interaction items; Step 1.2: Iterate through all users using the method in Step 1.1 to complete the process. ; Step 1.3: Create a hypergraph of user-project interactions. Mapped to user adjacency graph : Step 1.3.1: In the user-project interaction hypergraph, each hyperedge represents a user. The following criteria are used to determine whether the hyperedges in the user-project hypergraph are connected: in, It is a decision function; it returns 1 if there are items between the hyperedges, and 0 otherwise. and Representing users respectively and users The super-edge, Indicates intersection; Step 1.3.2: For For users, the following function is used to calculate the similarity between users: in, This represents a single interactive item in the intersection. The similarity score for a single public interaction item is represented by... Calculate the similarity between users and use it as the weight of the edges; Step 1.4: Connect the user adjacency graph As an initialization of the user's implicit relationship network; Step 2, Extraction Hypergraph base order Reconstruction The structure; Step 3: Fusing Hypergraph with Layered Graph Convolutional Networks and implicit relationship networks The feature information is used to obtain a higher-order representation of the user; Step 4: Disseminate user preference information on the knowledge graph; Step 5: Predict the probability of user interaction with the project.

2. The recommendation method based on hypergraph basalt optimization of multivariate user representation according to claim 1, characterized in that: Step 2 also includes the following steps: Step 2.1: Calculate the reliability of the hypergraph order based on its interaction patterns and number of interactions. in The reliability score represents the hypergraph motif. This indicates the total number of nodes connected by the hyperedge. Indicates the hypergraph basis order. This indicates whether there are common interactive items in the hypergraph's basic order; 1 indicates if they exist, and 0 otherwise. Step 2.1.1: Set a threshold and filter unreliable hypergraph motifs based on reliability scores; Step 2.2: Calculate similarity weights and establish directions based on multi-user interactions; Step 2.2.1: Since the two users interact with more than one hypergraph motif, normalization is performed based on all hypergraph motifs: in This refers to the similarity scoring based on binary users in Step 1.3.

2. It is a hypergraph sequence reliability score. It is a normalization function based on the attention mechanism, used to balance the values ​​between binary and ternary relations. End users and users Similarity score between them; Step 2.2.2: Select and establish directionality based on the weight of the relevant hypergraph motif types; Step 2.3, Traversal All edges, and in Find all relevant hypergraph base orders ; Step 2.4: Reconstruct according to Step 2.2 The structure.

3. The recommendation method based on hypergraph basalt optimization of multivariate user representation according to claim 1, characterized in that: Step 3 also includes the following steps: Step 3.1, As input to a hypergraph convolutional network, and leveraging the correlation between items to obtain the embedding of the user representation; Step 3.1.1, from the super edge The collection of connected nodes and user history interaction items To perform further propagation, its convolution operation is defined as: in, It is an activation function. Indicates the first Layer and first Trainable weight matrices between layers Indicates the relevant superedge, Represents the bias matrix; Step 3.1.2: Multiple hypergraph convolutional layers may lead to numerical instability, therefore symmetric normalization is added, and it is represented in matrix form as follows: in, It is an activation function. and It is the degree matrix of nodes and hyperedges in the hypergraph. Indicates to Transpose of; Step 3.1.3: Obtain the user embedding representation by aggregating the embeddings of items on the hyperedge: in, This represents the feature vector representation of the user. It is an activation function. Represents aggregate functions, It is the representation of the user's historical interaction items in the hypergraph; Step 3.2, Obtain user information Embedded representation in; Step 3.2.1: Obtain the user's embedding representation based on the hypergraph convolutional network and use it as the input to the graph convolutional network. The convolution operation is defined as follows: in Represents a node Neighbors It is the product of the square roots of the degree of the nodes. From node To the node scalar weights, It is an activation function; Step 3.3: Use the user representation output in the graph convolutional network as the hypergraph. For the input in the sequence, repeat steps 3.1 and 3.2 until the required number of hierarchical convolutional layers is met; Step 3.4: Obtain the user's final embedded representation .

4. The recommendation method based on hypergraph basalt optimization of multivariate user representation according to claim 1, characterized in that: Step 4 also includes the following steps: Step 4.1: To better understand users' knowledge preferences, a function is used to calculate a score between users and relationships, thus determining the importance of different relationships in the knowledge graph to users. Among them, A computed function representing a user's preference for relation types. This represents the user's historical set of interaction items. In the knowledge graph Neighborhood set, This indicates the relationship type when associating with a neighbor. Calculate users and The inner product, and These represent the feature vector representations of users and the feature vector representations of relationships, respectively. Step 4.2: Calculate the user's preference weight for the project: in, Indicates the number of user interaction items. express In Node degree, Calculate users For the project Preference weights; Step 4.3: Aggregate item representations in the knowledge graph based on the attention mechanism; Step 4.3.1: Based on the user's preference weight for the project, determine the ratio of the user's preference weight for the relationship. For projects with high interaction rates, aggregate neighbor entities more evenly. Step 4.3.2: Sample entity features from the project's local domain and aggregate the project's multi-hop neighbors to explore more information in the knowledge graph. The project aggregation is represented as follows: in, This represents the activation function. Indicates project The neighborhood group, Indicates the number of sampled neighbors. and These are the bias matrix and the bias terms, respectively. This represents the feature vector representation of the project.

5. The recommendation method based on hypergraph basalt optimization of multivariate user representation according to claim 1, characterized in that: Step 5 also includes the following steps: Step 5.1: Represent the project , and user representation The interaction probability is calculated as follows: in, This represents the predictive scoring function; Step 5.2: Iterate through all possible user item pairs using a negative sampling strategy. The loss function is as follows: in, yes - Regularization loss function, , and It is a balancing hyperparameter. It is cross-entropy loss. It is the user aggregation loss. It is the aggregate loss of the project. express Layered user-level convolutional aggregation, express Layer propagation of project convolutional aggregation, and These are the actual interaction and the predicted interaction probability, respectively. Step 5.3, Model Performance Evaluation: Step 5.3.1 Evaluation Metrics: For click-through rate prediction scenarios, accuracy (AUC), ACC, and F1 scores are used to evaluate model performance; for Top-K recommendation scenarios, Precision@K and Recall@K are used to evaluate model performance.

Images