A social user implicit relationship recommendation method and device

Abstract

The present application provides a social user implicit relationship recommendation method and device, the method comprises the following steps: obtaining user data of multiple social platforms; according to the user data, a plurality of groups of heterogeneous information graphs based on user social information are constructed; according to the heterogeneous information graph, clustering processing is carried out respectively through a double information based multi-view attribute graph clustering method, a rule based clustering method and an attention mechanism based graph neural network recommendation clustering method, three kinds of clustering results are obtained; the three kinds of clustering results are fused to obtain a target recommendation result, which can improve the efficiency of recommendation in the user social network.

Description

Technical Field

[0001] This article belongs to the field of computer technology, specifically relating to a method and apparatus for recommending implicit relationships among social users. Background Technology

[0002] Social networks are important platforms for communication. With the rapid development of the internet and information computing, a vast amount of information exists, with massive amounts generated every moment. Finding useful information from this massive amount of data is becoming increasingly difficult for users. Since everyone's interests are different, recommendation systems that can provide personalized content have emerged and become a hot topic. Recommendation systems analyze user behavior and recommend suitable information to users, satisfying their individual needs. They are designed to predict users' potential interests in items by learning embeddings.

[0003] Therefore, improving the efficiency of social recommendations has become an urgent technical problem to be solved. Summary of the Invention

[0004] To address the aforementioned problems in existing technologies, the purpose of this paper is to provide a method and apparatus for recommending implicit relationships among social users, which can improve the efficiency of social recommendations.

[0005] To solve the above-mentioned technical problems, the specific technical solution presented in this paper is as follows:

[0006] On the one hand, this paper provides a method for recommending implicit relationships among social users, the method comprising:

[0007] Acquire user data from multiple social media platforms;

[0008] Based on the user data, construct multiple sets of heterogeneous information graphs based on user social information;

[0009] Based on the heterogeneous information graph, clustering is performed using a dual-information-enhanced multi-view attribute graph clustering method, a rule-based clustering method, and an attention-based graph neural network recommendation clustering method, respectively, resulting in three clustering results.

[0010] The three clustering results are then fused to obtain the target recommendation result.

[0011] Furthermore, the step of constructing multiple sets of heterogeneous information graphs based on user social information according to the user data includes:

[0012] Based on the user data, node information and attribute information are extracted. The node information includes user information and type information, and the attribute information is used to represent the relationship between nodes.

[0013] Based on the node information and the attribute information, a heterogeneous information graph based on user social information is generated.

[0014] Furthermore, based on the heterogeneous information graph, clustering processing is performed using a dual-information enhanced multi-view attribute graph clustering method, including:

[0015] The graph editor module is used to extract features from each heterogeneous information graph to obtain a low-level representation of each graph, specifically as follows: Among them W l Let represent the parameters of the l-th layer, e represent the layer number, and σ represent the activation function. H represents the hidden representation. v Let X represent the low-level representation of the v-th view, where X is the attribute feature matrix and A is the low-level representation of the v-th view. b Represented as a normalized adjacency matrix;

[0016] Based on the low-level representation in each heterogeneous information graph, the maximum mutual information value in each heterogeneous information graph is calculated using the mutual information maximization module, where the maximum mutual information value represents the target potential high-level representation recovered in the heterogeneous information graph;

[0017] A specific information reconstruction module is used to reconstruct specific information for each heterogeneous information graph, so that the module that maximizes mutual information obtains specific information from each heterogeneous information graph. The specific information reconstruction module includes the following three-layer graph neural network: in It is the l-th hidden representation of the reconstructed v-th view, θ l This represents the parameter of the l-th layer, where r is the number of layers. It is the reconstruction-specific information matrix of the v-th view, used to reconstruct the features of the v-th view;

[0018] The clustering results of the dual-information enhanced multi-view attribute graph clustering method are obtained by clustering the maximum mutual information in each heterogeneous information graph through a self-supervised clustering module.

[0019] Furthermore, the maximum mutual information value in each heterogeneous information graph is calculated using the following formula:

[0020]

[0021] Among them, I(H v S) represents mutual information. It is H v In the j-th row, S is the potential high-level representation recovered from all views, S = MLP([H 1 H 2 ;…;H V MLP(·) is a subnetwork consisting of several fully connected layers, where S and H are maximized.v When there is mutual information between them, the mutual information loss function is defined as:

[0022] Furthermore, the clustering method includes K-means clustering;

[0023] The step of clustering the maximum mutual information values ​​in each heterogeneous information graph using a self-supervised clustering module to obtain the clustering results based on the dual-information enhanced multi-view attribute graph clustering method includes:

[0024] The maximum mutual information value in each heterogeneous information graph is clustered using the sketch clustering method of K-means clustering to obtain the clustering result based on the dual-information enhanced multi-view attribute graph clustering method.

[0025] Furthermore, based on the heterogeneous information graph, clustering is performed using a rule-based clustering method, including:

[0026] Rules are extracted from the user data in each heterogeneous information graph to obtain a set of association rules;

[0027] Pruning is applied to identify duplicate rules in the association rule set, and duplicate rules are removed through redundancy checks to obtain non-redundant association rules;

[0028] In association rules, the preceding term is defined as the preceding term and the following term as the following term. By extracting features from the preceding and following terms in non-redundant association rules, a rule feature vector composed of two feature vectors is obtained.

[0029] Principal component analysis is performed on the rule feature vectors to obtain the dimensionality-reduced rule feature vectors;

[0030] The DBSCAN algorithm is used to cluster the dimensionality-reduced regular feature vectors to obtain the clustering results of the rule-based clustering method.

[0031] Furthermore, based on the heterogeneous information graph, clustering is performed using a graph neural network recommendation clustering method based on an attention mechanism to obtain clustering results, including:

[0032] A restart-based random walk strategy is used to sample nodes in each heterogeneous information graph to obtain node embeddings for the nodes in the heterogeneous information graph. The node embeddings are low-level representations of the nodes, representing the structural attributes and semantic information of the nodes in the heterogeneous information graph.

[0033] Based on the heterogeneous information graph, two types of meta-paths are defined: user-project meta-path and user-user meta-path, with node representations as follows: Where, n t It is the current node, nt+1 It is the next node, and B is the node type. Indicates that B is a neighbor of v. t+1 The node of type ρ represents the specified meta-path rule;

[0034] Based on the node embedding, a hierarchical attention network is used to calculate the weights of nodes in the two meta-paths to obtain the weight coefficients of each node with respect to its neighboring nodes, expressed as follows: in, k represents the set of neighbors of node u based on the meta-path ρ. The importance of node u to node v in the metapath ρ;

[0035] Based on the weight coefficients of each node to its neighboring nodes and the node embeddings of each node's neighboring nodes, the preferences of each node are clustered to obtain the clustering results for each node.

[0036] Furthermore, the process of fusing the three clustering results to obtain the target recommendation result includes:

[0037] The three clustering results are compared pairwise to obtain three similarity results;

[0038] Determine whether at least one of the three similarity results exceeds a preset value;

[0039] If so, the two clustering results corresponding to similarity results exceeding the preset value will be used as the target recommendation results;

[0040] If not, the two clustering results corresponding to the maximum similarity result will be used as the target recommendation result.

[0041] On the other hand, this paper also provides a social user implicit relationship recommendation device, the device comprising:

[0042] The acquisition module is used to acquire user data from multiple social media platforms;

[0043] The construction module is used to construct multiple sets of heterogeneous information graphs based on the user data and user social information.

[0044] The clustering module is used to perform clustering processing on the heterogeneous information graph using a clustering method based on dual-information enhanced multi-view attribute graph, a rule-based clustering method, and a graph neural network recommendation clustering method based on attention mechanism, respectively, to obtain three clustering results;

[0045] The fusion module is used to fuse the three clustering results to obtain the target recommendation result.

[0046] Finally, this document also provides a recommended device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor, when executing the computer program, implements the method described above.

[0047] Using the above technical solution, this paper provides a method and apparatus for recommending implicit relationships among social users. The method includes: acquiring user data from multiple social platforms; constructing multiple sets of heterogeneous information graphs based on user social information based on the user data; performing clustering processing on the heterogeneous information graphs using a dual-information enhanced multi-view attribute graph clustering method, a rule-based clustering method, and an attention-based graph neural network recommendation clustering method to obtain three clustering results; and fusing the three clustering results to obtain the target recommendation result. This method can improve the efficiency of recommendations in user social networks.

[0048] To make the above and other objects, features and advantages of this document more apparent and understandable, preferred embodiments are described below in detail with reference to the accompanying drawings. Attached Figure Description

[0049] To more clearly illustrate the technical solutions in the embodiments or prior art described herein, the accompanying drawings used in the description of the embodiments or prior art will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this article. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0050] Figure 1 This document illustrates the steps of a method for recommending implicit relationships among social users, as provided in an embodiment.

[0051] Figure 2 This document illustrates a flowchart of a method for recommending implicit relationships among social users, as provided in an embodiment.

[0052] Figure 3 This document illustrates a flowchart of a clustering method within a social user implicit relationship recommendation method provided in this embodiment.

[0053] Figure 4 This document illustrates a flowchart of a clustering method within a social user implicit relationship recommendation method provided in this embodiment.

[0054] Figure 5 This document shows a schematic diagram of the structure of a social user implicit relationship recommendation device provided in an embodiment of the invention;

[0055] Figure 6 A schematic diagram of the recommended device provided in the embodiments of this article is shown.

[0056] Explanation of symbols in the attached drawings:

[0057] 510. Acquisition Module; 520. Construction Module; 530. Clustering Module; 540. Fusion Module. Detailed Implementation

[0058] The technical solutions in the embodiments described below will be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments described herein, and not all of the embodiments. Based on the embodiments described herein, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this document.

[0059] It should be noted that the terms "first," "second," etc., used in the specification, claims, and accompanying drawings herein are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, apparatus, product, or device that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or devices.

[0060] Social networks are important platforms for communication. With the rapid development of the internet and information computing, a vast amount of information exists, with massive amounts generated every moment. Finding useful information from this massive amount of data is becoming increasingly difficult for users. Since everyone's interests are different, recommendation systems that can provide personalized content have emerged and become a hot topic. Recommendation systems analyze user behavior and recommend suitable information to users, satisfying their individual needs. They are designed to predict users' potential interests in items by learning embeddings.

[0061] To improve the efficiency of information recommendation in social networks, this paper provides a method for recommending implicit relationships among social users. Figure 1 This is a schematic diagram illustrating the steps of a method for recommending implicit relationships among social users, as provided in the embodiments herein. This specification provides the operational steps of the method described in the embodiments or flowcharts, but based on conventional or non-creative labor, more or fewer operational steps may be included. The order of steps listed in the embodiments is merely one possible execution order among many and does not represent the only possible execution order. In actual system or device products, the methods shown in the embodiments or accompanying drawings can be executed sequentially or in parallel. Specifically, as shown in the embodiments or accompanying drawings... Figure 1As shown, the method may include:

[0062] S101: Obtain user data from multiple social media platforms;

[0063] S102: Based on the user data, construct multiple sets of heterogeneous information graphs based on user social information;

[0064] S103: Based on the heterogeneous information graph, clustering is performed using a dual-information-enhanced multi-view attribute graph clustering method, a rule-based clustering method, and an attention-based graph neural network recommendation clustering method, respectively, to obtain three clustering results;

[0065] S104: The three clustering results are fused together to obtain the target recommendation result.

[0066] This specification can be understood as follows: Based on the spatial characteristics of heterogeneous information graphs (i.e., heterogeneous graphs), graph processing is performed on heterogeneous graphs constructed from user social data in social networks. At the same time, three clustering methods are performed in parallel, and the results of the three clustering methods are further fused to obtain the final recommendation result. This paper can accurately and efficiently realize the recommendation efficiency and capability in users' social networks.

[0067] Furthermore, the step of constructing multiple sets of heterogeneous information graphs based on user social information according to the user data includes:

[0068] Based on the user data, node information and attribute information are extracted. The node information includes user information and type information, and the attribute information is used to represent the relationship between nodes.

[0069] Based on the node information and the attribute information, a heterogeneous information graph based on user social information is generated.

[0070] This can be understood as the user data being social data from various platforms, used to construct the required dataset, and then undergoing a series of operations to determine and predict the relationships between users. Furthermore, to effectively and uniformly model user information, type attribute information, and various related relationships, a heterogeneous information attribute graph is constructed, defined as G = (V, E). In the heterogeneous information graph, V represents multiple types of node sets, and E represents the relationships between nodes. |E|=M. Defined as a multi-view property graph. This represents the number of views. For example, nodes can be usernames, user IDs, tweets, or tweet topics, and attribute information can include follow relationships, tweet retweet relationships, tweet comment relationships, etc. Through the above process, a heterogeneous graph in a social network can be quickly constructed based on user data.

[0071] It's important to note that traditional homogeneous graphs contain only one type of node and edge. Therefore, when building a graph neural network, all nodes share the same model parameters and have the same feature space. In contrast, heterogeneous graphs can contain more than one type of node and edge, thus allowing different types of nodes to have different dimensions of features or attributes.

[0072] Heterogeneous graph neural networks have been proposed as a modeling method for complex information networks. A heterogeneous graph consists of many nodes and edges. Nodes in a heterogeneous graph include users and items, while edges consist of interactions between users and items. Because heterogeneous graph neural networks are very flexible in modeling heterogeneous data, they are often used to represent rich auxiliary information in recommendation systems. Through the representation of heterogeneous graph neural networks, recommendation systems can be viewed as similarity searches of meta-paths on heterogeneous graph neural networks.

[0073] In the embodiments of this specification, clustering processing is performed using a dual-information enhanced multi-view attribute graph clustering method, including:

[0074] The graph editor module is used to extract features from each heterogeneous information graph to obtain a low-level representation of each graph, specifically as follows: Among them W l Let represent the parameters of the l-th layer, e represent the layer number, and σ represent the activation function. H represents the hidden representation. v Let X represent the low-level representation of the v-th view, where X is the attribute feature matrix and A is the low-level representation of the v-th view. v Represented as a normalized adjacency matrix;

[0075] Based on the low-level representation in each heterogeneous information graph, the maximum mutual information value in each heterogeneous information graph is calculated using the mutual information maximization module, where the maximum mutual information value represents the target potential high-level representation recovered in the heterogeneous information graph;

[0076] A specific information reconstruction module is used to reconstruct specific information for each heterogeneous information graph, so that the module that maximizes mutual information obtains specific information from each heterogeneous information graph. The specific information reconstruction module includes the following three-layer graph neural network: in It is the l-th hidden representation of the reconstructed v-th view, θ l This represents the parameter of the l-th layer, where r is the number of layers. It is the reconstruction-specific information matrix of the v-th view, used to reconstruct the features of the v-th view;

[0077] The clustering results of the dual-information enhanced multi-view attribute graph clustering method are obtained by clustering the maximum mutual information in each heterogeneous information graph through a self-supervised clustering module.

[0078] The basic idea of ​​multi-view clustering can be understood as exploring the consensus and complementary information embedded in multi-view data and learning clustering partitions suitable for all views. Multi-view spectral clustering methods aim to learn soft clustering index matrices directly from the original data.

[0079] Information maximization (or contrastive learning) has achieved great success in computer vision and has inspired much research on unsupervised learning. It aims to maximize the consistency between the deep representation obtained by performing a GNN on the input map and the augmented map.

[0080] Furthermore, the maximum mutual information value in each heterogeneous information graph is calculated using the following formula:

[0081]

[0082] Among them, I(H v S) represents mutual information. It is H v In the j-th row, S is the potential high-level representation recovered from all views, S = MLP([H 1 H 2 ;…;H V MLP(·) is a subnetwork consisting of several fully connected layers, where S and H are maximized. v When there is mutual information between them, the mutual information loss function is defined as:

[0083] Optionally, the self-supervised clustering module includes K-means clustering;

[0084] The step of clustering the maximum mutual information values ​​in each heterogeneous information graph using a self-supervised clustering module to obtain the clustering results based on the dual-information enhanced multi-view attribute graph clustering method includes:

[0085] The maximum mutual information value in each heterogeneous information graph is clustered using the sketch clustering method of K-means clustering to obtain the clustering result based on the dual-information enhanced multi-view attribute graph clustering method.

[0086] In the embodiments of this specification, clustering processing is performed based on the heterogeneous information graph using a rule-based clustering method, including:

[0087] Rules are extracted from the user data in each heterogeneous information graph to obtain a set of association rules;

[0088] Pruning is applied to identify duplicate rules in the association rule set, and duplicate rules are removed through redundancy checks to obtain non-redundant association rules;

[0089] In association rules, the preceding term is defined as the preceding term and the following term as the following term. By extracting features from the preceding and following terms in non-redundant association rules, a rule feature vector composed of two feature vectors is obtained.

[0090] Principal component analysis is performed on the rule feature vectors to obtain the dimensionality-reduced rule feature vectors;

[0091] The DBSCAN algorithm is used to cluster the dimensionality-reduced regular feature vectors to obtain the clustering results of the rule-based clustering method.

[0092] This can be understood as everything having certain rules, and social users are no exception. Hidden associations between social users can be mined through rule-based clustering. First, data is collected and preprocessed using different preprocessing techniques, such as stemming, affixation, extended acronyms, and negation substitution. Then, the most relevant features are selected. Association rules are generated by considering the selected features and terms, for example, using Competitive Group Optimization (CSO) and Gravity Search Algorithm (GSA) to select the most relevant terms for rule generation. Afterward, duplicate rules are pruned. Furthermore, DBSCAN-based clustering is introduced to group non-duplicate rules into candidate rules to identify rules among social users. This paper obtains rules among users through their social data, and shared interests among users are accurately mined with the help of these rules, thereby uncovering hidden associations among social users.

[0093] In the embodiments of this specification, based on the heterogeneous information graph, clustering is performed using a graph neural network recommendation clustering method based on an attention mechanism to obtain clustering results, including:

[0094] A restart-based random walk strategy is used to sample nodes in each heterogeneous information graph to obtain node embeddings for the nodes in the heterogeneous information graph. The node embeddings are low-level representations of the nodes, representing the structural attributes and semantic information of the nodes in the heterogeneous information graph.

[0095] Based on the heterogeneous information graph, two types of meta-paths are defined: user-project meta-path and user-user meta-path, with node representations as follows: Where, n t It is the current node, n t+1 It is the next node, and B is the node type. Indicates that B is a neighbor of v. t+1 The node of type ρ represents the specified meta-path rule;

[0096] Based on the node embedding, a hierarchical attention network is used to calculate the weights of nodes in the two meta-paths to obtain the weight coefficients of each node with respect to its neighboring nodes, expressed as follows: in, k represents the set of neighbors of node u based on the meta-path ρ. The importance of node u to node v in the metapath ρ;

[0097] Based on the weight coefficients of each node to its neighboring nodes and the node embeddings of each node's neighboring nodes, the preferences of each node are clustered to obtain the clustering results for each node.

[0098] Network representation learning, also known as network embedding or graph embedding, aims to represent nodes in a network as low-dimensional, real-valued, and dense vectors. This results in vectors that can be represented and reasoned within a vector space, and can be easily and conveniently used as input to machine learning models. Consequently, the resulting vector representations can be applied to common applications in social networks, such as visualization, node classification, link prediction, and community detection. They can also be used as social edge information in other common tasks such as recommendation systems.

[0099] Network representation learning can extract rich network structure and semantic information from heterogeneous graphs, addressing the cold start and data sparsity issues of traditional recommendation algorithms. It has been widely applied to specific data mining tasks, using low-dimensional vectors to represent the structure and semantic information of heterogeneous graphs and maximizing the use of auxiliary information from nodes and edges. Furthermore, different users will have different preferences for different node attributes and different meta-paths. Specifically, different node attributes can have different impacts on recommendation tasks. Additionally, users can have different preferences for different meta-path-based information. Attention mechanisms are widely used in data mining tasks; therefore, hierarchical attention networks are designed to learn user preferences for neighboring nodes and meta-paths.

[0100] This specification describes an embodiment that effectively fuses multi-dimensional representation vectors related to users and social interactions. It learns user preferences for neighboring nodes and meta-paths through an attention network, thereby enabling recommendation and clustering. Combined with a clustering method based on a total variation graph neural network, it jointly determines the clustering results, ultimately improving the accuracy and efficiency of clustering.

[0101] In the embodiments of this specification, the process of fusing the three clustering results to obtain the target recommendation result includes:

[0102] The three clustering results are compared pairwise to obtain three similarity results;

[0103] Determine whether at least one of the three similarity results exceeds a preset value;

[0104] If so, the two clustering results corresponding to similarity results exceeding the preset value will be used as the target recommendation results;

[0105] If not, the two clustering results corresponding to the maximum similarity result will be used as the target recommendation result.

[0106] This can be understood as follows: since the results of a single clustering method may contain errors, this application runs three clustering methods in parallel and merges the three clustering results to jointly determine the final clustering result. Specifically, by comparing the closeness (i.e., similarity) of the three clustering results, the more similar the clustering results are, the closer they are to the true recommendation. This is based on the premise that the three clustering results have a certain degree of reliability. Therefore, the more similar the clustering results are, the more accurate they are. Otherwise, if the similarity difference is large, it indicates that at least one clustering result is misaligned. Therefore, we can select at least two clustering results with high similarity (e.g., exceeding a preset value, or, if the similarity does not exceed a preset value, select the clustering result with the highest similarity) as the target recommendation result. This can further improve the efficiency and accuracy of recommendations in social networks.

[0107] The method provided in this application can better uncover implicit connections between social users, and has the following effective effects:

[0108] 1. In order to effectively extract and utilize the rich structural and semantic information in heterogeneous graphs, a method of joint representation of multiple features is proposed. By using multi-order topological structure information in heterogeneous information networks, the representation of node features in sparse network connection cases is enhanced.

[0109] 2. To achieve better clustering, a Dual Information Augmented Multi-View Attribute Graph Clustering (DIAGC) method is proposed. This method introduces a Specific Information Reconstruction (SIR) module to separate consensus and specific information exploration from multiple views, enabling the graph neural network to capture more fundamental low-level representations. Furthermore, a Mutual Information Maximization (MIM) module maximizes the consistency between the underlying high-level and low-level representations, allowing the high-level representations to satisfy the desired clustering structure with the help of a Self-Supervised Clustering (SC) module.

[0110] 3. An attention-based graph neural network recommendation clustering algorithm is presented, which effectively fuses multi-dimensional representation vectors of users, social interactions, etc. The attention network learns user preferences for neighboring nodes and meta-paths, thereby performing recommendations and clustering. It is combined with a clustering method based on a total variation graph neural network to jointly determine the clustering results, aiming to improve the accuracy of clustering.

[0111] 4. To improve the clustering effect, a sketch solution method for accelerating k-means clustering is introduced, which aims to identify the optimal k-means cluster.

[0112] This specification also provides an embodiment of a method for recommending implicit relationships among social users, such as... Figure 2 The diagram shown is a flowchart of the method, which specifically includes the following steps:

[0113] 1. Data Collection

[0114] We acquire relevant user data from various social media platforms to construct the required dataset, and then perform a series of operations to determine and predict the relationships between users.

[0115] 2. Construction of Heterogeneous Diagrams

[0116] To effectively and uniformly model user information, type attribute information, and various related relationships, a heterogeneous information attribute graph is constructed, defined as G = (V, E). In the heterogeneous information graph, V represents multiple types of node sets, and E represents the relationships between nodes. |E|=M. Defined as a multi-view property graph. This represents the number of views.

[0117] 3. Attention-based Graph Neural Network Recommendation Clustering Algorithm

[0118] like Figure 3 The diagram shows a flowchart of a graph neural network recommendation clustering algorithm based on an attention mechanism. The recommendation algorithm can improve the recommendation quality of the recommendation model by effectively fusing multi-dimensional representation vectors such as user and social data. At the same time, it can also uncover implicit connections between social users to a certain extent.

[0119] 1) Joint representation of multiple features of information

[0120] Given a heterogeneous graph, we need to generate low-dimensional vector representations e for all nodes v in the graph, and the generated node embeddings need to include the structural properties and semantic information of the heterogeneous graph. We use network representation learning techniques to efficiently represent the heterogeneous graph and apply the obtained auxiliary information to subsequent recommendation methods.

[0121] Furthermore, a restart-based random walk strategy can be used to sample heterogeneous neighbor nodes. A restart-based random drift sampling strategy is used to sample neighbor nodes. Starting from the initial node v, a random walk iterates with probability P to the current node's neighbors or back to the initial node until the sampling strategy successfully collects a fixed number of nodes, denoted as R(v). The sampled neighbor nodes are classified by type. The top K nodes are selected based on their frequency of occurrence in R(v), and they are used as the neighbor set associated with a node V of type t.

[0122] 2) Metagraph Construction

[0123] In order to better discover metapaths and to embed the input graph into the metagraph, which can also help with clustering, the metagraph is constructed before defining the metapaths.

[0124] MG =<MV,MV,E,ME> , where MG represents the metagraph; V represents the vertex set of the graph; MV represents the vertex set of the metagraph; E represents the edge set of the metagraph; and ME represents the edge set of the metagraph.

[0125] Meta image segment: MG i ={ev j},ev j ∈(V∪MV∪E∪ME), where MG i Represented as a meta-image segment; ev j Elements that belong to the union of vertices, metavertices, edges, and metaedges.

[0126] Metagraph Vertex: mv i =<{atr k},MG j >,mv i ∈MV, where mv i The vertices of the metagraph belong to the metagraph vertex set MV; atr k Indicates attribute, MG j This is a meta-image segment.

[0127] 3) Hierarchical attention network

[0128] Two types of metapaths are defined here: the first is the user-project metapath, which reflects the interaction between users and projects; the second is the user-user metapath, which reflects the social relationships between users.

[0129]

[0130] Where n t It is the current node, n t+1 It is the next node, and B is the node type. Indicates that B is a neighbor of v. t+1The node of type ρ represents the specified metapath rule.

[0131] Different users will have different preferences for different node attribute information and different meta-paths. Specifically, different node attribute information can have different impacts on recommendation tasks. Furthermore, users can have different preferences for different meta-path-based information. Attention mechanisms have been widely used in data mining tasks; therefore, hierarchical attention networks are designed to learn users' preferences for neighboring nodes and meta-paths.

[0132] Attention mechanisms are used to learn the weights between nodes. Given a pair of nodes (u, v) connected by a meta-path ρ, node-level attention... Importance can be learned, which means the importance of node v to node u. The importance of node pairs based on metapaths can be expressed as:

[0133]

[0134] Among them atten node This represents a node-level attention network. The above equation shows that the weights of node pairs (u,v) based on metapaths depend on the characteristics of a given metapath ρ. It is asymmetric, meaning that the importance of node u to node v and the importance of node v to node u may not be the same, because asymmetry is an important property of heterogeneous graphs. The Sofimax function is used for normalization. The final weighted average is obtained through the following methods:

[0135]

[0136] Where k represents the set of neighbors of node u based on the meta-path ρ. The weight coefficients of (u,v) depend on their properties. Weight focus They are asymmetric, meaning they contribute differently to each other. This is not only due to the connection order in the molecule, but also because they have different neighbors.

[0137] The meta-path-based embedding of node u can be clustered using the projected features of its neighbors, with the following coefficient combinations:

[0138]

[0139] in It is the learned embedding of node u of the metapath ρ. Each node is aggregated through its neighbor embeddings.

[0140] For any user, the implicit vector representation of the user's relationship to a particular item is represented by its neighborhood items and ratings. Then, the corresponding weights are generated through an attention network, and finally, all implicit vectors are fused to obtain the implicit vector representation of the user in the interaction network.

[0141]

[0142] in In the interactive network, user u i The implicit vector representation of , where C represents all ratings given by users to the project, and aggre represents the fusion function.

[0143]

[0144] Where θ(·) is the active function. b are neural network parameters, α ij It is the weight of project j relative to user i.

[0145] The implicit vector representation of a user in a social network is represented by the fusion of vectors of the user's neighbors in the user's social network.

[0146]

[0147]

[0148] β iI It is the weight of user I to user i.

[0149] 4) Top-K ranking model learning

[0150] Embed e using the user's final node u As input to a multilayer fully connected layer with nonlinear transformations.

[0151] z u =ReLU(W L ReLU(W1e u +b1)+b L )

[0152] Among them W L and b L These represent the weight matrix and bias vector for each layer, respectively. User preferences for the project are as follows:

[0153]

[0154] Where w p and b p Let represent the weight matrix and the bias vector, respectively. Large likelihood estimation is used to model the optimization objective of the recommendation task.

[0155] L=∑(y u log(p u )+(1-y u log(1-p) u )+λ||θ|| 2)

[0156] Here, λ and θ are the parameter set and regularization coefficient of the model, respectively, which can be updated using stochastic gradient descent.

[0157] 4. A Dual-Information Enhanced Multi-View Attribute Graph Clustering (DIAGC) Method

[0158] like Figure 4 The diagram illustrates the flowchart of the Dual Information Augmented Multi-View Attribute Graph Clustering (DIAGC) method. To improve the effectiveness of mining user association rules, another Dual Information Augmented Multi-View Attribute Graph Clustering method is introduced. This method consists of four modules: a Graph Encoder (GE) module, a Mutual Information Maximization (MIM) module, a Specific Information Reconstruction (SIR) module, and a Self-Supervised Clustering (SC) module. Finally, a sketch solution method to accelerate k-means clustering is presented to identify the optimal k-means cluster.

[0159] 1) Graph Encoder (GE) Module: A graph encoder module is used to capture a low-level representation of each view. Specifically, the graph encoder is used as a non-linear feature extractor to mine deep non-linear correlations in the attribute graph embedded in each view.

[0160]

[0161]

[0162] Among them, W l Let represent the parameters of the l-th layer, e represent the layer number, and σ represent the activation function. H represents the hidden representation. v This represents the low-level representation of the v-th view. X is the attribute feature matrix, A v It is represented as a normalized adjacency matrix.

[0163] 2) Mutual Information Maximization (MIM) Module: Maximizing mutual information encourages similarity of the same samples across different views. To explore each H... v The embedded high-level consensus information utilizes the Mutual Information Maximization (MIM) module:

[0164]

[0165] in It is H v In the j-th row, S is the potential high-level representation recovered from all views.

[0166] S = MLP([H 1 H 2 ;…;H V ])

[0167] The MLP(·) is a subnetwork composed of several fully connected layers. When maximizing S and H... v When there is mutual information between them, instead of simply H v By combining these elements, S can capture the intrinsic information of each view. The mutual information loss function is defined as follows:

[0168]

[0169] 3) Specific Information Reconstruction (SIR) Module: Reconstructing specific information enables the MIM module to capture more important information suitable for all views. Specifically, the Specific Information Reconstruction module is implemented by a three-layer graph neural network:

[0170]

[0171]

[0172] in It is the l-th hidden representation of the reconstructed v-th view, θ l This represents the parameter of the l-th layer, where r is the number of layers. It is the reconstruction-specific information matrix of the v-th view, used to reconstruct the features of the v-th view.

[0173] 4) Self-Supervised Clustering (SC) Module: Since multi-view attribute graph clustering is an unsupervised task, ground truth labels are unavailable, which poses a challenge to network training. To address this issue, a self-supervised clustering module is employed to guide network training. Specifically, KL divergence is used:

[0174]

[0175] Where Q is the soft-label distribution, i.e., q ij The probability that the i-th instance belongs to the j-th cluster is measured, where k is the number of clusters and P is the target distribution. ij Calculated using the t-distribution:

[0176]

[0177] Where μ=[μ1;μ2;…;μ c Let ] be the cluster centroid, c be the number of clusters, and o be the degrees of freedom of the t-distribution. The target distribution P is defined as:

[0178]

[0179] Among them, the second q is used ijThis makes the target distribution P denser. It enhances the learning process of Q by pushing away samples belonging to different clusters and clustering samples of the same cluster together. To initialize the cluster centroids, k-means clustering is performed on S before network training, after which μ is adaptively updated in each training iteration. The SC module makes S a high-level cluster-oriented representation.

[0180] The mutual information maximization module ensures H v It includes common information about the views, while the specific information reconstruction module reconstructs the features of the v-th view. After T iterations, the final clustering result is obtained by performing k-means clustering on S to uncover implicit relationships among social users.

[0181] 5) A method for accelerating the sketching solution of k-means clustering

[0182] This module introduces a sketch-based method for k-means clustering using a semi-definite procedure (SDP) to identify the optimal k-means clustering. This method works well when the clusters are sufficiently separable. If the original clusters are well separated, the random subset satisfies the proximity condition with a high probability, which in turn means recovering the expected clusters within that subset. The entire dataset can then be partitioned based on which of these cluster centers is closest.

[0183] First, we introduce the necessary symbols to clarify the proximity conditions. For And Γ∈∏(n,k), for each in k is the number of clusters, and the index set [n] := {1,…,n} represents the number of point clusters. ∏(n,k) is the set of k non-empty sets that partition [n]. The tuple X := {x i} i∈[n] It is R d A point in the interval is defined as:

[0184]

[0185] Among them, c S Representation group It is R d Points and non-empty sets in The corresponding index, c T Representation group It is R d Points and non-empty sets in The corresponding index, Let represent a d×n matrix whose i-th column is equal to zero. Note that... Capture Bisect the center of mass and c T The distance to the hyperplane, and pass The size of T scales some variance measures across the entire dataset. Intuitively, k-means clustering is easier when the clusters separate well, for example, when the following quantities are positive:

[0186]

[0187] Given Considering the quantities of Γ∈∏(n,k):

[0188]

[0189] Theorem: There exists an explicit shape parameter C:(R d ) n The following rules apply to ×∏(n,k)→[0,∞]:

[0190] (a) Assumption And Γ∈∏(n,k) satisfy prox(X,Γ)>0 and Then C(X, Γ) < ∞;

[0191] (b)C With d, k and Directly related.

[0192] When clusters are more separated, shape parameters are smaller, therefore, according to the above theorem, sketch size can be smaller.

[0193] First, the behavior of the sketch-based solution method is considered as n→∞. It can scale according to the shape of the data. Whenever the data exhibits a sufficiently good shape, the method can recover the optimal k-means clustering.

[0194] Then, the lower bound for the optimal k-means clustering is provided, which can be estimated by the expected value of the random sketch Y, ESDP(Y,k).

[0195] The optimal clustering of k-means is identified through the steps described above.

[0196] 5. Rule-based clustering model framework

[0197] Everything has its rules, and social media users are no exception. Hidden connections between social media users can be uncovered through rule-based clustering. First, data is collected and preprocessed using various preprocessing techniques, such as stemming, affixation, expanded acronyms, and negation substitution. Then, the most relevant features are selected. Association rules are generated by considering the selected features and terms, for example, using Competitive Group Optimization (CSO) and Gravity Search Algorithms (GSA) to select the most relevant terms for rule generation. Afterward, duplicate rules are pruned. Furthermore, DBSCAN-based clustering is introduced to group non-duplicate rules into candidate rules to identify rules among social media users.

[0198] 1) Removal of duplicate rules

[0199] Pruning is applied to identify duplicate rules in the tweet database. Specifically, for a rule R, α→S, where α is a universally existing interest or α appears everywhere in R, is defined as follows: This represents the original association. In contrast, rules with these terms do not provide any new knowledge about the data R→S, and these rules are long and difficult to understand. These types of repetition rules filter based on their conditional probabilities. Therefore, a redundancy check process is needed to reduce duplicate rules.

[0200] 2) Feature Vector Creation

[0201] Once the non-redundant rules are obtained, they are defined as numerical feature vectors. This interpretation is based on determining the distances between rules and facilitates rule-based clustering. Association rules are described by two distinct sets of interest terms: one on the left called the antecedent and the other on the right called the result. Feature vectors are created by merging the features of the antecedent and result of a rule. All features associated with a unique item may exist on their respective sides.

[0202] consider It consists of the i-th preceding feature and the j-th resulting feature; k l k r This represents the length of the feature vectors for the corresponding antecedent and outcome. A new feature domain is considered in the range of -1 to 1, which defines the degree of association between the considered feature and the antecedent or outcome of the rule. A negative range implies contradiction, while a positive range implies similarity. In particular, all feature values ​​are represented as correlation scores between the feature and the most relevant term in the rule.

[0203] The correlation score between features and items is represented as a correlation coefficient. These coefficients are then expressed as similarity, or contradiction. This representation relies on the assumption that items appearing together may be identical, while items present in different social user data may contradict each other. To avoid complex correlations, ρ ≤ 0.05 is used, and another coefficient is specified as 0. In any relationship, a small correlation coefficient indicates a small correlation.

[0204] These feature vectors are compared to analyze how the rules are related to each other. This can be used to cluster the rules and find unwanted rules.

[0205] 3) Rule-based clustering

[0206] To identify rules among users, association rules are clustered, and implicit associations are validated. Principal Component Analysis (PCA) is applied to the feature vectors defining association rules before clustering. This minimizes the computational cost of clustering large rule sets. Based on the DBSCAN algorithm, Euclidean distance on numerical feature vectors is used for clustering. DBSCAN classifies rules into cores or outliers. Rules considered cores are deemed to indicate associations between social users. Outliers are candidate unpopular rules, i.e., rules without associations. To determine whether these outliers are unpopular, they are ultimately compared to cores through contradiction verification.

[0207] The difference between two user interest itemsets is estimated by determining the cosine similarity of their feature vectors. Itemsets are considered either different or similar when the cosine value is greater than a given threshold. The similarity threshold for the preceding feature vectors is denoted as ε1, and the conflict threshold for the subsequent feature vectors is denoted as ε2, where ε1 is a positive integer representing the existence of rules between users, and ε2 is a negative integer representing the existence of rules between users.

[0208] The DBSCAN algorithm for detecting rules among candidate users requires two factors: min Ps and ds. min Ps represents the minimum number of data points required to create a dense region, and ds represents the aspect where the highest distance between a data point and its neighbors is considered to belong to the same cluster. The highest dimension of contradictory clusters is M. τ =τ|N|rules, where N represents the set of all rules and τ represents a small subset. Then, ds gradually decreases until DBSCAN finds a cluster that does not break the conflict-free constraint. Therefore, rules among users are obtained from the data, and common interests among users are accurately mined with the help of these rules. This further allows for the discovery of implicit connections between social users.

[0209] This specification's embodiments utilize clustering methods to better uncover implicit connections between social users. Three clustering methods are employed to mine these connections, with the main benefits being as follows:

[0210] 1. Network representation learning can extract rich network structure and semantic information from heterogeneous graphs, which can solve the problems of cold start and data sparsity in traditional recommendation algorithms. It has been widely used in specific data mining tasks and uses low-dimensional vectors to represent the structure and semantic information of heterogeneous graphs. It also makes full use of the auxiliary information of nodes and edges in heterogeneous graphs, providing a new approach to improve the accuracy of recommendation models.

[0211] 2. A joint multi-feature representation of information in heterogeneous information networks is employed. For user / item interaction and user social domains, the representation of node features in sparse network connections is enhanced by using multi-order topological structure information in heterogeneous networks. In the case of sparse rating matrices, the recommendation quality of the recommendation model is improved through the effective fusion of multi-dimensional representation vectors of users, items, ratings, and social interactions.

[0212] 3. A dual-information-enhanced multi-view attribute graph clustering method is proposed, which cleverly separates consensus and specific information exploration. It fully leverages the nonlinearity of low-level representations to generate potential high-level clustering-oriented representations. It uses a mutual information maximization module and a specific information reconstruction module to mine rich information embedded in the deep representations of each view for use in the potential high-level representation. A specific information reconstruction module is also introduced, enabling the GCN encoder to extract purer consensus representations. A self-supervised clustering module forces the potential high-level representations to become clustering-oriented representations. By training these modules simultaneously, they can mutually reinforce each other to achieve better clustering results.

[0213] Based on the methods provided above, embodiments of this specification also provide a social user implicit relationship recommendation device, such as... Figure 5 As shown, the device includes:

[0214] Module 510 is used to acquire user data from multiple social media platforms;

[0215] The construction module 520 is used to construct multiple sets of heterogeneous information graphs based on the user data.

[0216] Clustering module 530 is used to perform clustering processing on the heterogeneous information graph using a clustering method based on dual-information enhanced multi-view attribute graph, a rule-based clustering method, and a graph neural network recommendation clustering method based on attention mechanism, respectively, to obtain three clustering results;

[0217] The fusion module 540 is used to fuse the three clustering results to obtain the target recommendation result.

[0218] The beneficial effects obtained by the above-described device are the same as those obtained by the above-described method, and will not be described in detail in the embodiments of this specification.

[0219] This embodiment provides a recommended device, the internal structure of which can be shown in the following diagram. Figure 6 As shown. The recommended device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When the computer program is executed by the processor, it implements a method for identifying the surface covering of the recommended device.

[0220] Those skilled in the art will understand that Figure 6 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the recommended device to which the present application is applied. Specific recommended devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0221] In one embodiment, a recommended device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above method embodiments.

[0222] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon that, when executed by a processor, implements the steps in the above method embodiments.

[0223] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0224] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0225] It should also be understood that, in the embodiments herein, the term "and / or" is merely a description of the relationship between associated objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this document generally indicates that the preceding and following associated objects have an "or" relationship.

[0226] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, computer software, or a combination of both. To clearly illustrate the interchangeability of hardware and software, the components and steps of the various examples have been generally described in terms of functionality in the foregoing description. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this document.

[0227] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0228] In the embodiments provided herein, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the couplings or direct couplings or communication connections shown or discussed may be indirect couplings or communication connections through some interfaces, devices, or units, or they may be electrical, mechanical, or other forms of connection.

[0229] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the embodiments described herein, depending on actual needs.

[0230] This document uses specific embodiments to illustrate the principles and implementation methods of this document. The descriptions of the embodiments above are only for the purpose of helping to understand the methods and core ideas of this document. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of this document. Therefore, the content of this specification should not be construed as a limitation of this document.

Claims

1. A method for recommending implicit relationships among social users, characterized in that, The method includes: Acquire user data from multiple social media platforms; Based on the user data, construct multiple sets of heterogeneous information graphs based on user social information; Based on the heterogeneous information graph, clustering is performed using a dual-information-enhanced multi-view attribute graph clustering method, a rule-based clustering method, and an attention-based graph neural network recommendation clustering method, respectively, resulting in three clustering results. The three clustering results are then fused together to obtain the target recommendation result; Based on the heterogeneous information graph, clustering is performed using a dual-information-enhanced multi-view attribute graph clustering method, including: The graph editor module is used to extract features from each heterogeneous information graph to obtain a low-level representation of each heterogeneous information graph. Based on the low-level representation in each heterogeneous information graph, the maximum mutual information value in each heterogeneous information graph is calculated using the mutual information maximization module, where the maximum mutual information value represents the target potential high-level representation recovered in the heterogeneous information graph; By using a specific information reconstruction module, specific information is reconstructed for each heterogeneous information graph, so that the module that obtains the maximum mutual information can obtain specific information in each heterogeneous information graph; The maximum mutual information in each heterogeneous information graph is clustered by a self-supervised clustering module to obtain the clustering results based on the dual-information enhanced multi-view attribute graph clustering method. Based on the heterogeneous information graph, clustering is performed using a rule-based clustering method, including: Rules are extracted from the user data in each heterogeneous information graph to obtain a set of association rules; Pruning is applied to identify duplicate rules in the association rule set, and duplicate rules are removed through redundancy checks to obtain non-redundant association rules; In association rules, the preceding term is defined as the preceding term and the following term as the following term. By extracting features from the preceding and following terms in non-redundant association rules, a rule feature vector composed of two feature vectors is obtained. Principal component analysis is performed on the rule feature vectors to obtain the dimensionality-reduced rule feature vectors; The DBSCAN algorithm is used to cluster the dimensionality-reduced regular feature vectors to obtain the clustering results of the rule-based clustering method.

2. The method according to claim 1, characterized in that, The step of constructing multiple sets of heterogeneous information graphs based on user social information according to the user data includes: Based on the user data, node information and attribute information are extracted. The node information includes user information and type information, and the attribute information is used to represent the relationship between nodes. Based on the node information and the attribute information, a heterogeneous information graph based on user social information is generated.

3. The method according to claim 1, characterized in that, The low-level representation of each heterogeneous information graph is as follows: ,in Let represent the parameters of the l-th layer, e represent the layer number, and σ represent the activation function. Indicates the first v The first view l Hidden representation of layers, Indicates the first v The hidden representation of the 0th layer of the view. Indicates the first v A low-level representation of a view. X For attribute feature matrix, Represented as a normalized adjacency matrix; The specific information reconstruction module includes the following three-layer graph neural network: ,in It is the first v The first reconstructed view l Layer hidden representation, It is the first v The 0th layer hidden representation of the reconstructed view. Indicates the first l Layer parameters, r It is the number of layers. It is the first v The reconstruction-specific information matrix of the first view is used to reconstruct the second view. v Features of each view.

4. The method according to claim 1, characterized in that, The maximum mutual information value of each heterogeneous information graph is calculated using the following formula: ； in, For mutual information, yes The j-th row, For the first v A low-level representation of a view. MLP(·) is a subnetwork consisting of several fully connected layers, when maximizing S and When there is mutual information between them, the mutual information loss function is defined as: .

5. The method according to claim 1, characterized in that, The self-supervised clustering module includes K-means clustering; The step of clustering the maximum mutual information values ​​in each heterogeneous information graph using a self-supervised clustering module to obtain the clustering results based on the dual-information enhanced multi-view attribute graph clustering method includes: The maximum mutual information value in each heterogeneous information graph is clustered using the sketch clustering method of K-means clustering to obtain the clustering result based on the dual-information enhanced multi-view attribute graph clustering method.

6. The method according to claim 1, characterized in that, Based on the heterogeneous information graph, a clustering process is performed using a graph neural network recommendation clustering method based on an attention mechanism to obtain the clustering results, including: A restart-based random walk strategy is used to sample nodes in each heterogeneous information graph to obtain node embeddings for the nodes in the heterogeneous information graph. The node embeddings are low-level representations of the nodes, representing the structural attributes and semantic information of the nodes in the heterogeneous information graph. Based on the heterogeneous information graph, two types of meta-paths are defined: user-project meta-path and user-user meta-path; the node representations are as follows: ,in, It is the current node. It is the next node. B For node type, express v Among the neighbors The node of type ρ represents the specified meta-path rule; Based on the node embedding, a hierarchical attention network is used to calculate the weights of nodes in the two meta-paths to obtain the weight coefficients of each node with respect to its neighboring nodes, expressed as follows: ;in, , k Represents a node based on the metapath ρ u The neighborhood group, Metapath ρ Middle node u For nodes v The importance of Represents a node u For nodes i Importance of the metapath ρ Represents a node u For nodes j Importance under the metapath ρ; Based on the weight coefficients of each node to its neighboring nodes and the node embeddings of each node's neighboring nodes, the preferences of each node are clustered to obtain the clustering results for each node.

7. The method according to claim 1, characterized in that, The process of fusing the three clustering results to obtain the target recommendation result includes: The three clustering results are compared pairwise to obtain three similarity results; Determine whether at least one of the three similarity results exceeds a preset value; If so, the two clustering results corresponding to similarity results exceeding the preset value will be used as the target recommendation results; If not, the two clustering results corresponding to the maximum similarity result will be used as the target recommendation result.

8. A social user implicit relationship recommendation device, characterized in that, The device includes: The acquisition module is used to acquire user data from multiple social media platforms. The construction module is used to construct multiple sets of heterogeneous information graphs based on the user data and user social information. The clustering module is used to perform clustering processing on the heterogeneous information graph using a clustering method based on dual-information enhanced multi-view attribute graph, a rule-based clustering method, and a graph neural network recommendation clustering method based on attention mechanism, respectively, to obtain three clustering results; The fusion module is used to fuse the three clustering results to obtain the target recommendation result; The clustering process, based on the heterogeneous information graph, is performed using a dual-information enhanced multi-view attribute graph clustering method, including: The graph editor module is used to extract features from each heterogeneous information graph to obtain a low-level representation of each heterogeneous information graph. Based on the low-level representation in each heterogeneous information graph, the maximum mutual information value in each heterogeneous information graph is calculated using the mutual information maximization module, where the maximum mutual information value represents the target potential high-level representation recovered in the heterogeneous information graph; By using a specific information reconstruction module, specific information is reconstructed for each heterogeneous information graph, so that the module that obtains the maximum mutual information can obtain specific information in each heterogeneous information graph; The maximum mutual information in each heterogeneous information graph is clustered by a self-supervised clustering module to obtain the clustering results based on the dual-information enhanced multi-view attribute graph clustering method. Based on the heterogeneous information graph, clustering is performed using a rule-based clustering method, including: Rules are extracted from the user data in each heterogeneous information graph to obtain a set of association rules; Pruning is applied to identify duplicate rules in the association rule set, and duplicate rules are removed through redundancy checks to obtain non-redundant association rules; In association rules, the preceding term is defined as the preceding term and the following term as the following term. By extracting features from the preceding and following terms in non-redundant association rules, a rule feature vector composed of two feature vectors is obtained. Principal component analysis is performed on the rule feature vectors to obtain the dimensionality-reduced rule feature vectors; The DBSCAN algorithm is used to cluster the dimensionality-reduced regular feature vectors to obtain the clustering results of the rule-based clustering method.

9. A recommended device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the method as described in any one of claims 1 to 7.

Images