Scientific research partner recommendation method based on attention mechanism and graph neural network

Abstract

The invention discloses a scientific research partner recommendation method based on an attention mechanism and a graph neural network, and provides an academic partner recommendation model AGNNRec based on the graph neural network and the attention mechanism. According to the method, a recommendation model is constructed based on a graph neural network, a context path is introduced to capture a high-order relationship between nodes, a multi-head relationship attention mechanism is introduced to learn embedded vectors of node high-order neighbors in parallel, and finally, recommendation of partners is completed by using a cosine similarity algorithm. Experiments are carried out on the model on a self-built data set and a public data set DBLP respectively, in the experiment using the DBLP data set, the effects of the model on Precision (at) 5 and NDGG (at) 5 are improved by 19.07% and 1.17% respectively, and the effectiveness of AGNNRec in science and technology partner recommendation tasks is shown.

Description

Technical Field

[0001] The present application relates to the technical field of scientific research collaborator recommendation methods, and in particular, to a scientific research collaborator recommendation method based on attention mechanism and graph neural network. Background Art

[0002] The existing academic network stores a huge amount of data, and it has become extremely difficult to find relevant collaborators for researchers from the huge amount of data. In recent years, with the vigorous development of science and technology, close connections between scientific and technological collaborators have become crucial. However, how to find suitable partners from the massive amount of academic literature and research data to achieve more efficient research cooperation and produce more influential scientific research results has become a major technical challenge. In this context, the science and technology service knowledge graph and its related science and technology collaborator recommendation have become a research field that has attracted much attention.

[0003] At present, many scholars at home and abroad have conducted theoretical research and experimental analysis on the recommendation methods of scientific and technological collaborators. Most of them mainly use the recommendation methods of scientific and technological collaborators in homogeneous networks. The proposed methods build networks through co-author relationships, colleague relationships and other relationships, and recommend scholars based on the topological structure. A random walk method is proposed, which performs truncated random walks on isomorphic graphs to generate node sequences similar to "sentences" and node embedding vectors for recommendation. Although the above method has certain interpretability, it cannot be applied to large academic knowledge graphs because only one relationship can be represented in homogeneous networks. Therefore, it is necessary to use heterogeneous information networks for collaborator recommendation. Compared with homogeneous information networks, heterogeneous information networks contain more semantics. It is more meaningful to use heterogeneous information networks for scientific and technological collaborator recommendation. Heterogeneous information networks incorporate more information into the network, such as researchers' personal information, citation relationships between researchers and their published papers, etc. The recommendation methods for scientific and technological collaborators for heterogeneous knowledge graphs mainly use network embedding methods based on meta-paths and random walk strategies. Among them, someone proposed a method based on meta-paths and attention mechanisms, and recommended academic collaborators based on node embedding. Someone proposed a random walk method based on topics for collaborator recommendation tasks. Someone proposed a collaborator recommendation method based on heterogeneous information networks. Chuan proposed a collaborator recommendation method based on meta-paths. The above methods are suitable for large academic knowledge graphs and have good recommendation effects. However, the main problem of the methods is that defining the original path requires strong expert domain knowledge. With the development of academic knowledge graphs, the difficulty of defining meta-paths will increase dramatically, increasing personnel costs; and due to the use of random walk strategies, semantic information will inevitably be lost. Inspired by neural networks, recently, the work of applying graph neural networks to recommendation systems has gradually increased, but most methods are applied to homogeneous networks. Graph neural networks are applied to YouTube recommendation systems, but there are challenges in applying graph neural networks to heterogeneous information network recommendations. Some people point out that there is a situation where long-distance information transmission fails when recommendations are applied to heterogeneous information networks.

[0004] With the rapid development of science and technology and the increasing closeness of academic cooperation, collaborator recommendation in the field of scientific research has become an important tool for improving research efficiency and promoting innovation. Among the existing scientific research collaborator recommendation methods, commonly used recommendation strategies include methods based on homogeneous information networks and methods based on heterogeneous information networks. However, these methods still have many problems and challenges when dealing with complex and diverse scientific research cooperation relationships:

[0005] Traditional recommendation methods based on homogeneous information networks regard network nodes as a single type and only consider direct relationships between researchers. This network structure is difficult to capture the complex associations between heterogeneous data, such as the multi-level relationships between researchers and their published papers, participating projects, and affiliated institutions. As a result, the semantic information of the recommendation system is relatively single and cannot accurately reflect the needs of cross-disciplinary and cross-field cooperation.

[0006] In some recommendation methods based on heterogeneous information networks, it is necessary to rely on meta-paths defined by experts to capture high-order relationships in the network. However, as the network scale expands and the structure becomes more complex, the definition of meta-paths becomes increasingly difficult, and the generalization ability of meta-paths in different scenarios is weak, which can easily lead to a decline in recommendation results.

[0007] In existing recommendation models based on random walks and network embedding, the semantic information between some nodes is easily ignored due to the limitations of path truncation and sampling strategies. In addition, these methods perform poorly when dealing with long-distance node relationships and cannot effectively integrate the complex relationships between nodes, resulting in inaccurate recommendation results of the model.

[0008] Most current collaborator recommendation models have low computational efficiency when dealing with large-scale heterogeneous information networks. Especially in scenarios involving multi-relationship and multi-path aggregation, the computational complexity of the model increases significantly, making it difficult to support real-time recommendation of large-scale academic data. Summary of the invention

[0009] This application provides a method for recommending scientific research collaborators based on attention mechanism and graph neural network, aiming to solve the problems of homogeneous information network limitations, meta-path dependence, semantic information loss and model optimization difficulties in the recommendation of scientific research cooperation relationships in the existing technology, as well as the need for efficient recommendation algorithms.

[0010] In the first aspect, a method for recommending scientific research collaborators based on an attention mechanism and a graph neural network is provided, the method comprising:

[0011] Construct a heterogeneous information network containing multiple node types and relationship types; the heterogeneous information network contains the following node types: researchers, papers, projects, patents, institutions, tasks and conferences; wherein edge types include but are not limited to: the relationship between researchers and paper writing, the relationship between papers and conference publishing, the relationship between researchers and patent writing, the relationship between researchers and project undertaking, and the relationship between projects and tasks;

[0012] Using a graph neural network model to initialize an embedding vector for each node in the heterogeneous information network and generate a corresponding context vector;

[0013] Introducing a semantic-aware attention mechanism to calculate the weights of the relationships between the target node and its neighbors of different node types;

[0014] Construct a relation-aware attention mechanism to quantify the contribution of different relation types in the graph;

[0015] Use multi-head relational attention mechanism to aggregate contextual neighbor information and improve model robustness;

[0016] Construct an optimization objective function to optimize the node embedding vector and context vector of each node in the network;

[0017] Calculate the cosine similarity based on the node vectors and generate a recommended list of scientific research collaborators.

[0018] In the above scheme, optionally, the node initialization process is completed by the following steps:

[0019] Generate a low-dimensional embedding vector for each node using a normal distribution;

[0020] Map the embedding vector to the initial context vector of the node to capture the neighbor semantic information of the node in different paths;

[0021] The node vector and context vector are input into the graph neural network model for subsequent training and optimization.

[0022] In the above solution, optionally, the implementation of the semantic-aware attention mechanism includes:

[0023] For each target node, the semantic weights of neighbor nodes are calculated according to different relationship types.

[0024] Concatenate the high-dimensional feature vectors of the node and its neighbors and map them to a single weight value through an activation function;

[0025] The softmax function is used to normalize the weights of all neighbor nodes, and information aggregation is completed through the attention coefficient.

[0026] In the above scheme, the optional implementation of the relation-aware attention mechanism includes:

[0027] Decompose the heterogeneous information network into multiple subgraphs, each subgraph corresponds to a node type;

[0028] Use node random discard mechanism to reduce the interference of noise nodes on the model;

[0029] Use the average aggregation method to integrate the node context information in the subgraph into a graph-level vector;

[0030] For each relationship edge containing the target node type, the weights of all relationship edges are normalized using the softmax function, and information aggregation is completed through the attention coefficient.

[0031] In the above scheme, optionally, the implementation of the multi-head relation attention mechanism includes:

[0032] Integrate the relational attention weights at different granularities, combine the semantic-aware attention and relation-aware attention into the final relational attention, and multiple heads calculate the contextual neighbor information of the target node in parallel according to the attention coefficients of the corresponding heads;

[0033] Concatenate the results of each head calculation to generate the final context vector of the node;

[0034] In the node embedding process, the node information is passed layer by layer, and the node vector is updated using weights at each layer.

[0035] In the above solution, optionally, calculating the cosine similarity based on the node vector includes:

[0036]

[0037] in, and hj are the final embedding vectors of nodes i and j respectively.

[0038] In the above solution, optionally, the step of generating the recommendation list includes:

[0039] Use cosine similarity to calculate the similarity between the target node and all other nodes of the same type;

[0040] Sort the similarities of all nodes and select the top-N nodes as recommended collaborators of the target node;

[0041] Use the update mechanism of heterogeneous networks to dynamically adjust node vectors and recommendation results.

[0042] In the second aspect, a scientific research collaborator recommendation system based on attention mechanism and graph neural network is provided, wherein the system comprises:

[0043] Storage module: used to store node and relationship data of heterogeneous information networks;

[0044] Processing module: used to perform information aggregation and node vector optimization of graph neural network models;

[0045] Output module: used to display the recommendation results of partners and provide users with query and selection functions.

[0046] In a third aspect, a computer device includes a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, the following steps are implemented:

[0047] Construct a heterogeneous information network containing multiple node types and relationship types; the heterogeneous information network contains the following node types: researchers, papers, projects, patents, institutions, tasks and conferences; wherein edge types include but are not limited to: the relationship between researchers and paper writing, the relationship between papers and conference publishing, the relationship between researchers and patent writing, the relationship between researchers and project undertaking, and the relationship between projects and tasks;

[0048] Using a graph neural network model to initialize an embedding vector for each node in the heterogeneous information network and generate a corresponding context vector;

[0049] Introducing a semantic-aware attention mechanism to calculate the weights of the relationships between the target node and different types of neighbor nodes;

[0050] Construct a relation-aware attention mechanism to quantify the contribution of different relation types in the graph;

[0051] Use multi-head relational attention mechanism to aggregate contextual neighbor information and improve model robustness;

[0052] Construct an optimization objective function to optimize the node embedding vector and context vector of each node in the network;

[0053] Calculate the cosine similarity based on the node vectors and generate a recommended list of scientific research collaborators.

[0054] In a fourth aspect, a computer program product comprises a computer program / instruction, wherein when the computer program / instruction is executed by a processor, the following steps are implemented:

[0055] Construct a heterogeneous information network containing multiple node types and relationship types; the heterogeneous information network contains the following node types: researchers, papers, projects, patents, institutions, tasks and conferences; wherein edge types include but are not limited to: the relationship between researchers and paper writing, the relationship between papers and conference publishing, the relationship between researchers and patent writing, the relationship between researchers and project undertaking, and the relationship between projects and tasks;

[0056] Using a graph neural network model to initialize an embedding vector for each node in the heterogeneous information network and generate a corresponding context vector;

[0057] Introducing a semantic-aware attention mechanism to calculate the weights of the relationships between the target node and different types of neighbor nodes;

[0058] Construct a relation-aware attention mechanism to quantify the contribution of different relation types in the graph;

[0059] Use multi-head relational attention mechanism to aggregate contextual neighbor information and improve model robustness;

[0060] Construct an optimization objective function to optimize the node embedding vector and context vector of each node in the network;

[0061] Calculate the cosine similarity based on the node vectors and generate a recommended list of scientific research collaborators.

[0062] Compared with the prior art, this application has at least the following beneficial effects:

[0063] Based on further analysis and research of existing technical problems, this application recognizes that the existing technology has the problems of homogeneous information network limitations, meta-path dependence, semantic information loss and model optimization difficulties, as well as the need for efficient recommendation algorithms in the recommendation of scientific research cooperation relationships. This application constructs a heterogeneous information network containing multiple node types and relationship types; the heterogeneous information network contains the following node types: scientific researchers, papers, projects, patents, institutions, tasks and conferences; wherein edge types include but are not limited to: the writing relationship between scientific researchers and papers, the publishing relationship between papers and conferences, the writing relationship between scientific researchers and patents, the undertaking relationship between scientific researchers and projects, and the attribution relationship between projects and tasks; a graph neural network model is used to initially identify each node in the heterogeneous information network. Initialize the embedding vector and generate the corresponding context vector; introduce the semantic-aware attention mechanism to calculate the weight of the relationship between the target node and different types of neighbor nodes; construct a relationship-aware attention mechanism to quantify the contribution of different relationship types in the graph; use the multi-head relationship attention mechanism to aggregate contextual neighbor information and improve the robustness of the model; construct an optimization objective function to optimize the node embedding vector and context vector of each node in the network; calculate the cosine similarity based on the node vector, and generate a recommendation list of scientific research collaborators. The method proposed in this application not only solves the shortcomings of homogeneous information networks in semantic capture, but also makes up for the defects of semantic information loss in existing heterogeneous information networks through the automatic learning of context paths and the introduction of multi-head relationship attention mechanisms. In addition, the present invention provides an efficient node information aggregation and recommendation strategy, which improves the accuracy and computational efficiency of scientific research collaborator recommendations, and can be widely used in interdisciplinary scientific research cooperation, technology project team building, and patent cooperation matching.

[0064] This application proposes an academic collaborator recommendation model AGNNRec (Attention Graph Neural Networks Recommendation) based on graph neural networks and attention mechanisms. This method builds a recommendation model based on graph neural networks, captures high-order relationships between nodes by introducing context paths, and introduces a multi-head relationship attention mechanism to learn the embedding vectors of high-order neighbors of nodes, and finally completes the recommendation of collaborators by using the cosine similarity algorithm. The model was experimented on a self-built dataset and a public dataset DBLP. In the experiment using the DBLP dataset, the model's performance on Precision@5 and NDGG@5 improved by 19.07% and 1.17% respectively, demonstrating the effectiveness of AGNNRec in the task of recommending scientific and technological collaborators.

[0065] This application proposes a method for recommending scientific and technological collaborators based on academic heterogeneous information networks, AGNNRec. The method used in this application is mainly based on the idea of ​​graph neural network to build the model, introduce context path and attention mechanism, and capture the implicit relationship between nodes by learning the embedding vectors of node context neighbors, so as to achieve the recommendation of collaborators. Compared with other methods, this method has the following innovations:

[0066] In order to solve the problem that existing methods use less semantic information for recommendation using homogeneous information networks, we propose to use the rich structural information of heterogeneous information networks, including the complex relationships between various entity types such as scientific and technological personnel, papers, projects, and fields, so as to more comprehensively explore the potential connections between collaborators.

[0067] In order to address the problem in existing methods that the model effect is overly dependent on expert domain knowledge due to the use of predefined meta-paths, we abandon the use of predefined meta-paths and propose a learning method using graph neural networks to directly learn relevant node representations from academic networks, making the model more generalizable and adaptable.

[0068] To address the problem in existing methods that graph neural networks lose too much semantic information when transmitting information over long distances, we introduced two relational attention mechanisms at different granularities and a path length attention mechanism to further improve the performance and effect of the recommendation model.

[0069] Through a large number of experiments and comparative analysis on real academic knowledge graph data, the effectiveness and advantages of the proposed method are verified, and its potential and application prospects in the task of recommending scientific and technological collaborators are proved. The research results of this embodiment have important theoretical significance and practical value for promoting cooperation and communication in scientific research and promoting the sharing and application of academic achievements. BRIEF DESCRIPTION OF THE DRAWINGS

[0070] Figure 1 A schematic diagram of a process for recommending scientific research collaborators based on an attention mechanism and a graph neural network provided in one embodiment of the present application;

[0071] Figure 2 A context path neighbor diagram provided for one embodiment of the present application;

[0072] Figure 3 A schematic diagram of the overall framework of a model provided for one embodiment of the present application;

[0073] Figure 4 A schematic diagram of the outline information of a DBLP data set provided for one embodiment of the present application;

[0074] Figure 5 A schematic diagram of the outline information of a self-built data set provided in one embodiment of the present application;

[0075] Figure 6 A schematic diagram of context path attention score provided for one embodiment of the present application;

[0076] Figure 7 A schematic diagram of model performance for different context path lengths provided for one embodiment of the present application;

[0077] Figure 8 A schematic diagram showing the recommended effects provided by an embodiment of the present application;

[0078] Fig. 9 FIG. 4 is a diagram showing the internal structure of a computer device in one embodiment. DETAILED DESCRIPTION

[0079] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0080] In one embodiment, Figure 1 As shown in the figure, a method for recommending scientific research collaborators based on attention mechanism and graph neural network is provided, which includes the following steps:

[0081] Construct a heterogeneous information network containing multiple node types and relationship types; the heterogeneous information network contains the following node types: researchers, papers, projects, patents, institutions, tasks and conferences; wherein edge types include but are not limited to: the relationship between researchers and paper writing, the relationship between papers and conference publishing, the relationship between researchers and patent writing, the relationship between researchers and project undertaking, and the relationship between projects and tasks;

[0082] Using a graph neural network model to initialize an embedding vector for each node in the heterogeneous information network and generate a corresponding context vector;

[0083] Introducing a semantic-aware attention mechanism to calculate the weights of the relationships between the target node and different types of neighbor nodes;

[0084] Construct a relation-aware attention mechanism to quantify the contribution of different relation types in the graph;

[0085] Use multi-head relational attention mechanism to aggregate contextual neighbor information and improve model robustness;

[0086] Construct an optimization objective function to optimize the node embedding vector and context vector of each node in the network;

[0087] Calculate the cosine similarity based on the node vectors and generate a recommended list of scientific research collaborators.

[0088] In this embodiment, the node initialization process is completed by the following steps:

[0089] Generate a low-dimensional embedding vector for each node using a normal distribution;

[0090] Map the embedding vector to the initial context vector of the node to capture the neighbor semantic information of the node in different paths;

[0091] The node vector and context vector are input into the graph neural network model for subsequent training and optimization.

[0092] In this embodiment, the implementation of the semantic-aware attention mechanism includes:

[0093] For each target node, the semantic weights of neighbor nodes are calculated according to different relationship types.

[0094] Concatenate the high-dimensional feature vectors of the node and its neighbors and map them to a single weight value through an activation function;

[0095] The softmax function is used to normalize the weights of all neighbor nodes, and information aggregation is completed through the attention coefficient.

[0096] In this embodiment, the implementation steps of the relation-aware attention mechanism include:

[0097] Decompose the heterogeneous information network into multiple subgraphs, each subgraph corresponds to a node type;

[0098] Use node random discard mechanism to reduce the interference of noise nodes on the model;

[0099] The node context information in the subgraph is integrated into a graph-level vector using the average aggregation method.

[0100] In this embodiment, the implementation of the multi-head relation attention mechanism includes:

[0101] Integrate the relational attention weights at different granularities, combine the semantic-aware attention and relation-aware attention into the final relational attention, and multiple heads calculate the contextual neighbor information of the target node in parallel according to the attention coefficients of the corresponding heads;

[0102] Concatenate the results of each head calculation to generate the final context vector of the node;

[0103] In the node embedding process, the node information is passed layer by layer, and the node vector is updated using weights at each layer.

[0104] In this embodiment, the calculation of cosine similarity based on node vectors includes:

[0105]

[0106] in, and hj are the final vector information of nodes i and j respectively.

[0107] In this embodiment, the step of generating the recommendation list includes:

[0108] Use cosine similarity to calculate the similarity between the target node and all other nodes;

[0109] Sort the similarities of all nodes and select the top-N nodes as recommended collaborators of the target node;

[0110] Use the update mechanism of heterogeneous networks to dynamically adjust node relationships and recommendation results.

[0111] Supports the construction of heterogeneous information networks and fully captures complex relationship information;

[0112] This embodiment constructs a heterogeneous information network containing multiple node types (such as researchers, papers, projects, patents, institutions and tasks) and the complex relationships between them. This network structure can not only capture the multi-level relationships in academic activities, but also support interdisciplinary cooperation between different entities, thus solving the problem that traditional homogeneous networks cannot capture complex semantic information.

[0113] This embodiment uses a graph neural network (GNN) model to replace the traditional meta-path definition, and captures the high-order relationship between nodes through automatic learning of context paths. This innovation avoids the limitations of meta-path construction that relies on expert knowledge, so that the model has stronger generalization ability and can adapt to scientific research cooperation recommendation tasks in different fields. Through the semantic-aware attention mechanism and the relationship-aware attention mechanism, the relationship types and semantic information of different nodes are weighted. The concept of context path is introduced to reduce the loss of semantic information between long-distance nodes, thereby improving the accuracy of the recommendation results. The context neighbor information is aggregated in parallel through the multi-head relationship attention mechanism to improve the robustness of the model. By transferring and updating the embedding vectors of the nodes layer by layer, and using negative sample node pairs to optimize the model, the robustness of the training is improved. Combined with cosine similarity to calculate the similarity between nodes, the present invention supports real-time recommendations in large-scale heterogeneous networks, solving the problem of low computational efficiency of existing recommendation systems in big data scenarios. Through the accurate capture of node relationships and multi-head information aggregation, a Top-N collaborator recommendation list can be generated for scientific researchers, avoiding recommendation errors caused by information loss in traditional models. This recommendation strategy is particularly effective in interdisciplinary and cross-institutional scientific research collaboration, and can promote the improvement of scientific research output and innovation capabilities.

[0114] This embodiment has significant technical effects in solving the problems of insufficient semantic information in traditional homogeneous networks, reliance on expert knowledge for meta-path construction, semantic information loss, and low computational efficiency. By constructing a heterogeneous information network, introducing contextual path learning and multi-head relationship attention mechanism, the present invention provides an efficient, accurate, and real-time scientific research collaborator recommendation method, which can be widely used in interdisciplinary cooperation, technology project team building, and patent cooperation matching, and provides better cooperation suggestions for scientific researchers.

[0115] In one embodiment, this embodiment proposes a method for recommending scientific and technological collaborators based on academic heterogeneous information networks, AGNNRec. The method used in this embodiment is mainly based on the idea of ​​graph neural network to build the model, introduce context path and attention mechanism, and capture the implicit relationship between nodes by learning the embedding vector of node context neighbors, so as to achieve the recommendation of collaborators. Compared with other methods, this method has the following innovations:

[0116] In order to solve the problem that existing methods use less semantic information for recommendation using homogeneous information networks, we propose to use the rich structural information of heterogeneous information networks, including the complex relationships between various entity types such as scientific and technological personnel, papers, projects, and fields, so as to more comprehensively explore the potential connections between collaborators.

[0117] In order to address the problem in existing methods that the model effect is overly dependent on expert domain knowledge due to the use of predefined meta-paths, this embodiment abandons the use of predefined meta-paths and proposes a learning method through graph neural networks to directly learn relevant node representations from academic networks, so that the model has stronger generalization ability and adaptability.

[0118] In order to address the problem in existing methods that graph neural networks lose too much semantic information when transmitting information over long distances, this embodiment also introduces two relational attention mechanisms at different granularities and a context path length attention mechanism to further improve the performance and effect of the recommendation model.

[0119] Through a large number of experiments and comparative analysis on real academic knowledge graph data, this example verifies the effectiveness and advantages of the proposed method, and proves its potential and application prospects in the task of recommending scientific and technological collaborators. The research results of this example have important theoretical significance and practical value for promoting cooperation and communication in scientific research and promoting the sharing and application of academic achievements.

[0120] The semantic-based collaborator recommendation method mainly uses the content information of scholars' publications for modeling and analysis, such as keywords, themes, abstracts, etc. Then, the interest preferences of the corresponding scholars are expressed through the publication information, and collaborators are recommended by calculating the similarity between scholars. Word embedding modeling for publications is widely used in collaborator recommendation models. For example, word2vec is used for collaborator recommendation. The word2vec model represents the information of proposals and reviewers and is used for reviewers to recommend scientific proposals. Topic modeling for publications is also widely used for collaborator recommendation. The LDA (Latent Dirichlet Allocation) model is one of them. LDA can reflect the relationship between publications and topics. For example, based on the LDA topic model, the relationship between the topics of scholars' publications is deeply mined to recommend collaborators. The collaborator recommendation system is applied. Based on topic modeling, the research interests of researchers are extracted from documents, the scholar vectors are learned, and the recommendation list is generated. Someone proposed a heterogeneous network embedding recommendation model HNERec. This method uses the LDA model for topic modeling, and uses the skip-gram model to embed nodes. Finally, a recommendation list is generated based on the similarity between the corresponding node vectors. Some people have combined the topic clustering model and the random walk model to explore the academic characteristics of researchers in different fields and provide suggestions for collaborators. However, since the above method does not build an information network, the semantic information available to each scholar is relatively fixed, and the extracted information has a strong timeliness, so the model effect is poor.

[0121] The collaborator recommendation methods based on topological structure mainly rely on the analysis of the topological structure in the knowledge graph. These methods usually use graph theory and network analysis techniques to extract features from the topological structure of the graph and use these features to recommend collaborators. For example, Benchettara et al. used a topological binary supervised machine learning method to solve the link prediction problem in the collaborative authoring network. Koh and Dobbie proposed an academic collaborator recommendation method that uses a co-author network and a weighted association rule method using a weighting mechanism called sociality. The prior art proposed a restart-based random walk method for co-editor recommendation in heterogeneous bibliographic networks. They used the form of metapaths to represent heterogeneous networks and refined edge weights for recommendations. The prior art proposed a network embedding model based on metapaths, first using the skip-gram model to embed nodes, then using the attention mechanism to aggregate node information and path information, and finally generating a recommendation list based on the similarity between nodes. The prior art proposed a metapath2vec model. It proposed two methods metapath2vec and metapath2vec++. The metapath2vec model extracts the metapath of nodes through random walks to connect each node to form a heterogeneous network, which is convenient for node vectorization. On this basis, metapath2vec++ improves the semantic relevance of each node. It can be applied to various scenarios, such as node classification, node clustering, similarity search, etc. Although the above methods have constructed a network, the way to obtain information for nodes is mostly sampling, which greatly reduces the amount of semantic information available for nodes.

[0122] The method proposed in this embodiment constructs an academic network based on the existing science and technology service knowledge graph constructed based on the LDA topic model, and does not adopt a sampling strategy. Instead, it uses a graph neural network to allow each person to learn embedded information in parallel on the network, and the learned information is more comprehensive.

[0123] Heterogeneous graphs are also called heterogeneous information networks, which are defined as And two mapping functions, node mapping function and relational mapping functions in For each There is a node type For every edge e∈ξ, there exists an edge type

[0124] The purpose of the recommendation system in this embodiment is to recommend nodes of the same type with high similarity to the target node. In this embodiment, the anchor nodes are all the target nodes in the recommendation system.

[0125] The main graph is a homogeneous subgraph of the type of the anchor node in the heterogeneous graph. Assuming the anchor node type is Author, the main graph is defined as For each node v∈V Author , each edge e∈ The auxiliary graph is an isomorphic subgraph of the type of non-anchor nodes in the heterogeneous graph. It is similar to the definition of the main graph. Assuming the type is P, the auxiliary graph is defined as For each node v∈V P , each edge Assume a heterogeneous graph Then there is 1 main graph and n-1 auxiliary graphs.

[0126] Given a heterogeneous graph, there exists a path connecting two vertices in the main graph, defined as {v i ,R K ,v j}, where R K is the context path, which is the path connecting two nodes. The path length is K, that is, the path contains K nodes. R K It can also be called a context node sequence. If such a context path exists, the present embodiment calls the two nodes context neighbors. For ease of understanding, an example is given as Figure 2 As shown, there are three context paths of different lengths between technician i and technician j, and i and j are context neighbors.

[0127] The symbols used in this embodiment are defined in the following table and explained:

[0128]

[0129]

[0130] This example proposes a collaborator recommendation model based on graph neural network and attention mechanism. The overall framework of the model is as follows: Figure 3 shown.

[0131] Figure 3 The overall framework of AGNNRec consists of three parts: Model information aggregation: Aggregate information for the target node. Before aggregating information, calculate the relationship attention for different relationship types; Node information optimization: Construct the objective function to optimize the node vector and context vector of each node in the network; Collaborator recommendation: Recommend scientific and technological collaborators for scientific and technological personnel

[0132] This embodiment first initializes the node to a low-dimensional vector representation H according to the different node types in the academic network. The initialized low-dimensional vector obeys the normal distribution, and the node context vector C0 is initialized with the node vector H. The context information vector is used to aggregate the information of all context neighbor nodes under the context paths of different lengths of the target node. Then the attention coefficient of the relationship is calculated by the context vector of the node, and finally the node vector and the relationship attention coefficient are used to aggregate the node information using the graph neural network model to update the node vector information. Among them, for each different relationship type connected to the target node, this embodiment designs two attention mechanisms of different granularity to calculate the importance of the corresponding relationship, which provides stronger interpretability for the aggregated information, and more scientifically updates the node information of the target node, greatly improving the performance of the following scientific and technological collaborator recommendation task.

[0133] Semantic-aware attention is mainly for node-level relationship attention. Because for each node, when passing information to the target node, the contribution made by each relationship type is different; for each target node, first aggregate neighbor nodes under the same relationship type. In order to distinguish the importance of each relationship type to the target node, this embodiment uses an attention mechanism inspired by GAT (Graph Attention Network). Calculate the attention coefficient of each relationship type connected to the target node to provide more information support for the subsequent graph neural network aggregation information.

[0134]

[0135]

[0136] in Used to map the concatenated high-dimensional features to a real number. is the semantic-aware attention weight of node i under relation type r of head h at layer l.

[0137] Relation-aware attention is mainly aimed at graph-level relationship attention. Because the contribution of each relationship type in the entire graph is different, the main purpose of relationship-aware attention is to calculate the attention scores of corresponding relationship types with different source node types and the same target node types in all relationship types, and provide more information support for subsequent graph neural network aggregation information. The original academic heterogeneous information network is decomposed into multiple subgraphs according to the node type. This embodiment uses a graph encoder to encode all node context vectors of each node type subgraph into a graph vector. In order to improve the robustness of the model, first, this embodiment uses a node dropout mechanism to randomly set part of the node information to 0, and then uses an average aggregation operation to aggregate all node context vectors in the subgraph into a graph-level representation vector for calculating the attention weight of each relationship type, and finally calculates the attention weight of each relationship type.

[0138] C p =nodeDropout(C P )(2.1)

[0139] g P =mean(C P )(2.2)

[0140] Among them, C P It is the context vector of all nodes in the subgraph of node type P. The nodeDropout method is used to randomly set the node vector to 0, and the mean method is an average aggregation operation.

[0141]

[0142] The graphEncoder method consists of formulas 2.1 and 2.2, which are used to obtain the graph vector representation of the subgraph. S is the source node type in the relationship type, r T is the target node type in the relationship type, and is the graph representation vector of the source node type subgraph and the target node type subgraph at layer l, where and is the context vector of all nodes in the hth head of all source node type subgraphs and all target node type subgraphs. This embodiment uses KLinear and Qlinear to map the source and target graph vectors into query vectors and key vectors. is the relation-aware attention weight score of the h-th head in the i-th layer for the corresponding relation type.

[0143] This embodiment combines semantic-aware attention and relation-aware attention into the final relational attention, integrates relational attention weights at different granularities, and provides a comprehensive weight for each node to aggregate neighbor node information under different relations. This embodiment uses a balance coefficient to balance the two attention mechanisms. μ∈(0,1];

[0144]

[0145] Finally, in order to facilitate node learning of network structure, this embodiment uses a graph neural network to aggregate the information of neighbor nodes under each relationship type in a weighted manner to update the target node vector information, and then concatenates the node vectors under different heads for the final node representation. The following formula assumes that the number of heads is 8.

[0146]

[0147] The above embodiment introduces the concept of context path, and this embodiment introduces the concept of context path probability. Specifically, <v i ,R K ,v j > represents that nodes i and j are connected by a context path of length K. The probability that i and j are connected in a context path of length K is the context path probability, and the formula is as follows:

[0148] p(v j |R K ,v i ,θ)(6)

[0149] Among them, θ is the parameter for calculating probability; according to the business demand training model, this embodiment selects the scholar node as the anchor node, and the objective function is to calculate the subgraph G author The co-occurrence probability of all nodes and their corresponding context neighbors in context paths of different lengths is shown as follows:

[0150]

[0151] in, That is, the neighbor nodes of node i in the k-length path, P k is a set of paths of length k connecting nodes i and j, α k It represents the attention weight when the path length is k.

[0152] Intuitively, since P k , this embodiment attempts to use the Breadth-First Search (BFS) algorithm to find all paths of length k between nodes, but as the length of the context path k increases, it is difficult to find all the context paths P for each node.k It has become an almost impossible task. In today’s booming graph neural network (GNN, GraphNeuralNetwork), inspired by the information transmission of graph neural network, this embodiment attempts to use the context vector c i ,c j To represent the information of the context neighbors of nodes i, j, according to the required context path length, the context neighbor information of the source node and the target node at the corresponding length is obtained by calling the graph neural network. In the process of calculating the probability, the vector information of the node is combined. In this embodiment, the loss function of this embodiment is changed as follows:

[0153]

[0154] Among them, ⊙ is the Hadamard product, and δ is the sigmoid function.

[0155] At the business level, the task in this embodiment is a recommendation task. In order to improve the robustness of the model, it is considered to add negative sample node pairs to the model training process, and the loss function is changed to the following:

[0156]

[0157] in is the non-neighbor node of node i in the context path of length k.

[0158] Research collaborator recommendation is a technology that recommends potential partners to researchers by analyzing their information. It is of great significance in the field of scientific research and can promote interdisciplinary cooperation and improve scientific research output and innovation capabilities.

[0159] In order to complete the scientific research collaborator recommendation task, this embodiment uses the node vectors of scholars obtained by the training model, uses the node vectors to calculate the cosine similarity between scholars to represent the possibility of scientific research cooperation between two scholars, and recommends based on the similarity. The calculation formula is:

[0160]

[0161] According to formula 10, this embodiment can obtain the similarity between node i and all nodes, and obtain the top-N collaborator names with the highest similarity by sorting according to the name of the target scholar.

[0162] Finally, for the content of this chapter, this embodiment is summarized as follows: This chapter and this embodiment introduces the overall framework of the model and some technical implementation details. In the training stage of this model, the node vector information and context vector information of the scholar are first initialized using a method that obeys the standard normal distribution. Then, the neighbor vectors of each scholar's node are aggregated mainly through the graph neural network model and the relational attention weight to obtain the context vector of the target scholar, which is combined with the node embedding vector of the target scholar himself to obtain the final embedding vector of the target scholar. Finally, the objective function is used to optimize the scholar's node vector, and the embedding vector information learned by the model is used to predict whether there is a context path between scholar pairs to optimize the node vector. After the training is completed, the scholar's node vector is used to recommend the Top-N scientific and technological collaborators for each scientific and technological personnel. The overall algorithm of this embodiment is as follows:

[0163]

[0164] This embodiment introduces the data set, evaluation criteria, baseline model, and experimental settings used in the experiment. Then, this embodiment compares the model AGNNRec of this embodiment with other models to prove the effectiveness of the model of this embodiment in the Top-N recommendation task.

[0165] The experimental environment of this embodiment is as follows: Intel(R) Core(TM) i7-8750H CPU@2.20GHz, memory size is 16GB, operating system is Windows 10 Professional 64-bit, experimental tool is PyCharm 2022.3.2, programming environment is Python 3.7. And use CUDA version 11.8 for GPU acceleration, the GPU model used is NVIDIA A800, the capacity is 80GB.

[0166] In order to verify the effectiveness of the AGNNRec model, this example uses the public dataset DBLPV10 (https: / / aminer.org / citation) as the experimental dataset. The DBLP dataset contains four types of nodes: author, paper, conference, and field. There are eight relationship edges between the four types of nodes: author->paper (written), paper->author (written), conference->paper (published), paper->conference (published), field->paper (included), paper->field (included). The outline information of the dataset is as follows: Figure 4 The detailed statistical information of the dataset is shown in Table 1.

[0167] At the same time, in order to verify the effectiveness of the model, this example extracts six types of data from the self-built technology service library to construct an academic knowledge graph, including: researchers, papers, patents, institutions, projects, and tasks; the outline information of the data set is as follows: Figure 5The detailed statistical information of the dataset is shown in Table 2.

[0168] Table 1 Statistics of DBLP dataset

[0169]

[0170] Table 2 Statistics of self-built dataset

[0171]

[0172] This embodiment first processes the DBLP data set. This embodiment stipulates that two scholars who jointly participate in publishing more than three papers in the data set are called academic collaborators. After extracting the record information of the author and the corresponding collaborator, this embodiment divides the record information into a training set and a test set with a ratio of 8:2. The data of the training set is used to train the model. For each collaborator in the test set, this embodiment randomly extracts 100 negative samples, puts them into the model, and uses the model to sort the collaborators. The model ranks the positive sample collaborator as the first, which is considered a successful collaborator prediction. In the experiment, the batch size of this embodiment is set to 256, the learning rate is set to 0.001, and the optimizer is Adam.

[0173] Then this embodiment processes the self-built data set. This embodiment stipulates that the joint participation in the publication of papers, patents, and scientific and technological projects in the data set is counted as one collaboration, and two scholars with a total of more than three collaborations are called academic collaborators. After extracting the record information of the authors and the corresponding collaborators, this embodiment divides the record information into a training set and a test set with a ratio of 8:2. The data of the training set is used to train the model. For each collaborator in the test set, this embodiment randomly extracts 100 negative samples, puts them into the model, and uses the model to sort the collaborators. If the model ranks the positive sample collaborator as the first, it is considered a successful prediction of the collaborator.

[0174] In order to verify the effectiveness of the model, the verification indicators used in this embodiment are Precision@k and NDGG@k. These two evaluation indicators are widely used in Top-N recommendation tasks. Precision refers to the proportion of truly relevant collaborators in the recommendation list to the total number of recommendations. NDCG (Normalized Discounted Cumulative Gain) takes into account the position of the collaborator in the recommendation list. The higher the position of the correctly recommended relevant collaborator, the higher the score.

[0175]

[0176] Among them, r i The correlation between the i-th result and the real result in the recommendation list is set to 0 / 1. Z represents the normalization constant, and its formula is as follows:

[0177]

[0178] This embodiment uses a public data set to analyze the context path length of the model and find the optimal hyperparameter of the path length for the model. Figure 6 The attention scores under different context path lengths are shown. This embodiment finds that no matter how the path length changes, the attention weight of the short path is always greater than that of the long path. As the length increases, the attention weight under the corresponding length decreases accordingly, which is consistent with common sense. This embodiment also finds that when the path length is greater than 4, the path attention scores of (k=4, k=5) are basically the same, indicating that the newly added path learns less relevant information, which has limited help in improving the performance of the model and increases the model calculation time. These high-order relationships should be discarded.

[0179] Figure 7 The effect of the model proposed in this embodiment when setting different context path lengths is shown. According to the experimental results, this embodiment finds that when the context path length is less than 5, the model effect is significantly improved, and when the context path length is greater than 4, the model effect is slightly reduced. Figure 6 The conclusions are consistent, further proving that in the collaborator recommendation task under the DBLP dataset, these high-order relationships with context paths greater than 4 should not be included in the calculation. For the model proposed in this embodiment, after experimental analysis, this embodiment selects k=4 as the best hyperparameter for the experiments in the following two sections.

[0180] In order to evaluate the effectiveness of the recommendation algorithm, this embodiment uses the DBLP dataset and selects 6 recommendation algorithms for comparative experimental analysis. In the comparative experiment, UserCF, MF, FMG, HNRWalker and CARec belong to the collaborator recommendation algorithm based on topological structure, among which UserCF, MF and FMG belong to the traditional collaborator recommendation model. HNERec belongs to the content-based collaborator recommendation method.

[0181] UserCF: This is a traditional user-based collaborative filtering recommendation model.

[0182] MF: This is a standard matrix factorization based recommendation model.

[0183] FMG: FMG is a recommendation model based on heterogeneous information networks. The original loss function is modified to pairwise ranking loss so that it can be applied to Top-N recommendation methods.

[0184] HNERec: This is a recommendation model based on heterogeneous information network embedding, relying on LDA topic modeling.

[0185] HNRWalker: This is a collaborator recommendation model based on the improved random walk algorithm.

[0186] CARec: This is a recommendation method based on heterogeneous information network embedding, relying on attention mechanism and meta-path

[0187] AGNNRec: This is the academic collaborator recommendation model based on graph neural network and attention mechanism proposed in this example.

[0188] Table 3 shows the experimental results of the model proposed in this embodiment and the experimental comparison model on the DBLP dataset for the Top-N recommendation task. The results are as follows:

[0189] Table 3 Comparative experimental results

[0190]

[0191] According to the experimental results, this embodiment is summarized as follows:

[0192] This example found that traditional recommendation methods such as UserCF and MF methods have poor effects in partner recommendation tasks. The FMG model uses matrix factorization (MF) + factorization machine (FM) method to solve the information fusion problem. The biggest difference between this method and the previous two methods is that it uses a heterogeneous information network, decomposes the heterogeneous information network into multiple meta-graphs, and recommends feature merging based on the similarity generated by each meta-graph. The Precision@5 of FMG is 0.1832, which is 42.34% higher than the Precision of MF. Experimental results show that in the recommendation system task, the fusion of heterogeneous information can express richer semantic information.

[0193] Among the methods based on topic modeling, the Precision@5 of HNERec is 0.2296, which is 25.32% higher than that of FMG. Among the methods based on random walk and network embedding, the Precision@5 of HNRWalker and CARec are 0.2368 and 0.2805, which are 29.25% and 53.11% higher than that of FMG, respectively. HNRWalker mainly uses the random walk method to generate meta-paths and merge meta-paths to improve recommendation performance; CARec uses the random walk method to generate meta-paths, uses the skip-gram model and CNN model to integrate semantic information and path information, and generates node embedding to improve recommendation performance.

[0194] The AGNNRec model proposed in this embodiment abandons the method of pre-defining meta-paths and uses a graph neural network model to autonomously learn high-order neighbor node information on a knowledge graph with rich patterns. Compared with the best performing model CARec in the comparative experiment, the model improved the evaluation indicators Precision@5 and NDGG@5 by 19.07% and 1.17% respectively.

[0195] Table 4 shows the experimental results of the model proposed in this embodiment on the Top-N recommendation task on the self-built dataset. The results are as follows:

[0196] Table 4 Experimental results of self-built dataset

[0197]

[0198] According to the experimental results, this embodiment found that the AGNNRec model proposed in this embodiment performed better on the self-built dataset. This is mainly because the heterogeneous information network constructed by the self-built dataset has more node types and a larger data volume than the heterogeneous information network constructed by the DBLP dataset, and the dataset contains more semantic information.

[0199] In order to further verify the effectiveness of each component of the model, this embodiment conducted an ablation experiment on the data set. The experimental results are shown in Table 5. The following explanation is made:

[0200] AGNNRec nosem It represents the model proposed in this embodiment without semantic perception attention

[0201] AGNNRec norel It represents the model proposed in this embodiment without relational awareness attention

[0202] AGNNRec is the model proposed in this embodiment.

[0203] Table 5 Ablation experiment results

[0204]

[0205] Based on the experimental results, this embodiment makes the following summary:

[0206] This embodiment proposes two model variants based on the proposed model: AGNNRec nosem It is a variant of the model, that is, the model without the semantic-aware attention mechanism, because the results under the two evaluation indicators show that the effect of the AGNNRec model is better than the variant model AGNNRec nosem , proving that using semantic-aware attention mechanism when aggregating information in the model can help improve the performance of the model.

[0207] The second variant is AGNNRec norel , that is, the model that removes the relationship-dependent attention mechanism, because the results under the two evaluation indicators show that the effect of the AGNNRec model is better than the variant model AGNNRec norel ,This embodiment can also draw the following conclusion: it is very necessary to use the relationship-aware attention mechanism in the information aggregation process, which helps to improve the model performance. According to the experimental results, this embodiment found that the relationship attention mechanism at the two granularities complement each other, providing more scientific auxiliary data for the target node to integrate adjacent information, which greatly improves the recommendation performance.

[0208] In order to facilitate the display of model recommendation results and verify the application effect of the model in actual projects, this embodiment uses a self-built data set for case analysis. This embodiment selects three researchers from the test set to demonstrate the effect of scientific and technological collaborators’ recommendation. The display results are as follows: Figure 8 As shown; finally, this embodiment illustrates the result of the partner recommendation through the effect diagram, and the result is shown in Table 6.

[0209] Table 6 Recommendation results

[0210]

[0211] According to the recommendation results, it is found that the recommendation effect of the model is consistent with the experimental results, and the relevant collaborators with successful recommendations are ranked higher.

[0212] In one embodiment, a scientific research collaborator recommendation system based on an attention mechanism and a graph neural network is provided, including:

[0213] Storage module: used to store node and relationship data of heterogeneous information networks;

[0214] Processing module: used to perform information aggregation and node vector optimization of graph neural network models;

[0215] Output module: used to display the recommendation results of partners and provide users with query and selection functions.

[0216] The specific implementation content of each module can be found in the above definition of the scientific research collaborator recommendation method based on attention mechanism and graph neural network, which will not be repeated here.

[0217] In one embodiment, a computer device is provided. The computer device may be a server, and its internal structure diagram may be as follows: Fig. 9As shown. The computer device includes a processor, a memory and a network interface connected through a system bus. The processor of the computer device is used to provide computing and control capabilities, and the network interface is used to communicate with an external terminal through a network connection. The computer device loads and runs a computer program to implement the above-mentioned scientific research collaborator recommendation method based on the attention mechanism and graph neural network.

[0218] Those skilled in the art will understand that Fig. 9 The structure shown in the figure is only a block diagram of a part of the structure related to the solution of the present application, and does not constitute a limitation on the computer device to which the solution of the present application is applied. The specific computer device may include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0219] In one embodiment, a computer program product is also provided, including a computer program / instruction, which, when executed by a processor, involves all or part of the process in the above-mentioned embodiment method.

[0220] The technical features of the above embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

Claims

1. A scientific research collaborator recommendation method based on attention mechanism and graph neural network, characterized in that: The method comprises: Construct a heterogeneous information network containing multiple node types and relationship types; the heterogeneous information network contains the following node types: researchers, papers, projects, patents, institutions, tasks and conferences; wherein the edge types include: the relationship between researchers and paper writing, the relationship between papers and conference publishing, the relationship between researchers and patent writing, the relationship between researchers and project undertaking, and the relationship between projects and tasks; Using a graph neural network model to initialize an embedding vector for each node in the heterogeneous information network and generate a corresponding context vector; Introducing a semantic-aware attention mechanism to calculate the weights of the relationships between the target node and neighboring nodes of different node types; Construct a relation-aware attention mechanism to quantify the contribution of different relation types in the graph; Use multi-head relational attention mechanism to aggregate contextual neighbor information and improve model robustness; Construct an optimization objective function to optimize the node vector and context vector of each node in the network; Calculate the cosine similarity based on the node vectors and generate a recommended list of scientific research collaborators.

2. The method according to claim 1, characterized in that The node initialization process is completed through the following steps: Generate a low-dimensional embedding vector for each node using a normal distribution; Map the embedding vector to the initial context vector of the node to capture the neighbor semantic information of the node in different paths; The node vector and context vector are input into the graph neural network model for subsequent training and optimization.

3. The method according to claim 1, characterized in that The implementation of the semantic-aware attention mechanism includes: For each target node, the semantic weights of neighbor nodes are calculated according to different relationship types. Concatenate the high-dimensional feature vectors of the node and its neighbors and map them to a single weight value through an activation function; The softmax function is used to normalize the weights of all neighbor nodes, and information aggregation is completed through the attention coefficient.

4. The method according to claim 1, characterized in that: The implementation steps of the relation-aware attention mechanism include: Decompose the heterogeneous information network into multiple subgraphs, each subgraph corresponds to a node type; Use node random discard mechanism to reduce the interference of noise nodes on the model; Use the average aggregation method to integrate the node context information in the subgraph into a graph-level vector; For each relationship edge containing the target node type, the weights of all relationship edges are normalized using the softmax function, and information aggregation is completed through the attention coefficient.

5. The method according to claim 1, characterized in that The implementation of the multi-head relation attention mechanism includes: Integrate the relational attention weights at different granularities, combine the semantic-aware attention and relation-aware attention into the final relational attention, and multiple heads calculate the contextual neighbor information of the target node in parallel according to the attention coefficients of the corresponding heads; Concatenate the results of each head calculation to generate the final context vector of the node; In the node embedding process, the node information is passed layer by layer, and the node vector is updated using weights at each layer.

6. The method according to claim 1, characterized in that The calculating of cosine similarity based on node vectors includes: Among them, hi and hj are the final vectors of nodes i and j respectively.

7. The method according to claim 1, characterized in that The step of generating the recommendation list includes: Use cosine similarity to calculate the similarity between the target node and all other nodes of the same type; Sort the similarities of all nodes and select the top-N nodes as recommended collaborators of the target node; Use the update mechanism of heterogeneous networks to dynamically adjust node vectors and recommendation results.

8. A system for implementing the scientific research collaborator recommendation method according to claims 1 to 7, characterized in that include: Storage module: used to store node and relationship data of heterogeneous information networks; Processing module: used to perform information aggregation and node vector optimization of graph neural network models; Output module: used to display the recommendation results of partners and provide users with query and selection functions.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the method according to any one of claims 1 to 7 are implemented.

10. A computer program product comprising a computer program / instructions, characterized in that When the computer program / instructions are executed by a processor, the steps of the method according to claims 1-7 are implemented.

Images