Academic cooperation relationship prediction method based on heterogeneous graph neural network

Abstract

The present invention discloses a method for predicting academic cooperation relationship based on heterogeneous graph neural network, including: collecting the author's work information to construct a work data set and a knowledge heterogeneous graph, wherein the nodes in the graph include author nodes, work nodes, belonging nodes and content nodes; extracting the information in the knowledge heterogeneous graph using DeepWalk algorithm and Text‑CNN model to obtain the embedding vector of the author node; aggregating the meta-paths in the knowledge heterogeneous graph based on the author nodes using a multi-head attention mechanism to obtain the long-term interest embedding representation of each author node; obtaining the author's short-term interest embedding representation based on LSTM and attention mechanism; constructing a binary classification model using a binary classification prediction method, and inputting the work data set into the binary classification model for training to obtain an academic cooperation prediction model. The present invention can capture the author's potential cooperation interest within a certain period of time, obtain a better recommendation effect, and improve the prediction accuracy.

Description

Technical Field

[0001] The present invention belongs to the technical field of information retrieval, and in particular relates to an academic cooperation relationship prediction method based on heterogeneous graph neural network. Background Art

[0002] In recent years, with the rapid development of science and technology, the trend of scientific research tasks and innovations shifting from relying on individual achievements to collaborative completion has become increasingly obvious. More than 90% of innovations in this century come from collaboration. At the same time, the development of science and technology has also brought about the diversification and complexity of research problems, and the cross-integration of multiple disciplines has become more and more common. As a result, cooperation is no longer as rare and single as it used to be, but has become more and more complex and diverse. The changing research directions of scholars and the increase in cooperative relationships have made it a very big challenge to find the most valuable scientific research collaborators from the huge academic digital libraries and networks. The same problem also exists in technical cooperation, which makes the prediction of academic / technical cooperation relationships more and more important.

[0003] However, existing academic / technical partnership predictions mainly adopt the following methods:

[0004] 1. Similarity-based models: Most similarity-based methods mainly use methods based on text similarity and structural similarity. Text similarity methods are generally based on content. Academic collaboration is recommended by collecting the similarity of semantic features of node attributes in the academic network, including scholar nodes and article nodes. For example, Tang, Wu, etc. developed a topic learning model CTL to recommend academic collaboration. The model uses LDA (Latent Dirichlet Allocation) to model the topic distribution of papers, and then integrates the restart random walk algorithm for academic recommendation. However, the calculation of text similarity depends on the relevant attribute information of the author. In many cases, we actually cannot obtain the attribute information of the author. On the other hand, structural similarity generally uses the topological information of nodes in the website to determine the similarity between authors. Starting from the structural information of the academic network, most early studies construct academic collaboration prediction models from topological features such as common neighbors, Jaccard similarity sparsity, etc., and usually use binary algorithms such as logistic regression to achieve this. For example, Zhou, etc. improved the random walk algorithm, added the influence of cooperation time frequency and neighbor node type, and realized the recommendation of academic cooperation relationships, but structural similarity has certain problems, such as the inability to solve the cold start problem.

[0005] 2. Hybrid-based methods: One type of method is to mix the structural and content features of the academic network. Some studies use different vertex proximity algorithms to calculate the structural similarity of scholar nodes, and extract the semantic similarity of paper nodes through vocabulary similarity algorithms, query likelihood language models, etc. to fuse structural similarity and content similarity, thereby obtaining the comprehensive similarity of scholar nodes. Scholars with higher similarity are more likely to establish collaborative relationships. These hybrid methods accumulate a lot of information, but depend largely on the choice of features.

[0006] 3. Model based on machine learning: This model mainly uses classification models, especially binary classification, to predict the cooperation between authors. Qi, Zhao, etc. regard the prediction of scientific research cooperation as a binary classification problem, and use structural similarity indicators as features in the co-authorship network to train the model. They also use logistic regression and support vector machines (SVM) as prediction models to predict the links between author nodes.

[0007] The above three methods are all based on homogeneous networks. Although the calculation is simple, they lose rich semantic information and may cause errors in the prediction results. In addition, there are also a small number of studies based on heterogeneous networks. Among them, Sun, Rick, etc. first applied meta-path-based topological features to heterogeneous networks and used logistic regression models to predict cooperative relationships. However, the computational complexity of several topological features used in this method is relatively high. In recent years, with the popularity of graph neural network algorithms, because academic networks are a typical heterogeneous graph structure, related algorithms of heterogeneous graph neural networks have also been more applied to academic networks and have achieved good performance, such as GAT (Graph Attention Networks) model, HAN (Heterogeneous graph attention network, heterogeneous graph attention network) model, etc. However, most of the existing heterogeneous graph neural network algorithms are general models that focus on the characteristics of the heterogeneous graph itself, but lack attention to the subdivision problem of academic cooperation prediction in heterogeneous academic networks. Summary of the invention

[0008] In view of the above problems, the present invention proposes a method for predicting academic cooperation relationships based on heterogeneous graph neural networks. To solve the above technical problems, the technical solutions adopted by the present invention are as follows:

[0009] A method for predicting academic cooperation relationships based on heterogeneous graph neural networks includes the following steps:

[0010] S1, collect the author's work information to build a work dataset, and build a knowledge heterogeneous graph based on the work dataset;

[0011] The knowledge heterogeneous graph is represented by G, G = (V, E), V represents the node set in the knowledge heterogeneous graph, E represents the edge set formed between the nodes in the knowledge heterogeneous graph, and the nodes in the node set V include author nodes, work nodes, attribution nodes and content nodes;

[0012] S2, using the DeepWalk algorithm and the Text-CNN model to extract the information in the knowledge heterogeneous graph established in step S1 to obtain the embedding vector of each author node;

[0013] S3, uses a multi-head attention mechanism to aggregate meta-paths in the knowledge heterogeneous graph based on author nodes to obtain the long-term interest embedding representation of each author node;

[0014] S4, based on LSTM and attention mechanism, obtains short-term interest embedding representation that characterizes the author's short-term cooperation tendency;

[0015] S5, using the binary classification prediction method to build a binary classification model, the work dataset is input into the binary classification model for training to obtain the academic cooperation prediction model.

[0016] The step S2 comprises the following steps:

[0017] S2.1, using word embedding to vectorize the information of each work node and content node into a corresponding text vector;

[0018] S2.2, input the text vectors obtained in step S2.1 into the Text-CNN model for feature vector extraction, and obtain the title feature vector and text feature vector of each author;

[0019] S2.3, using BiLSTM or a fully connected layer to fuse the title feature vector and the text feature vector obtained in step S2.2 to obtain the interest feature vector of each author;

[0020] S2.4, treating the knowledge heterogeneous graph established in step S1 as homogeneous, and using the DeepWalk algorithm to pre-train the knowledge heterogeneous graph to obtain a vector representation reflecting the topological relationship of each author node in the knowledge heterogeneous graph;

[0021] S2.5, fuse the interest feature vector obtained in step S2.3 and the vector representation obtained in step S2.4 to obtain the embedding vector of each author node.

[0022] The step S3 comprises the following steps:

[0023] S3.1, select meta-paths with length less than N from the knowledge heterogeneous graph based on the author node;

[0024] The meta-path includes a first meta-path indicating that there is a cooperative relationship between the author and other authors, a second meta-path indicating that there is a citation relationship between the author and other authors, a third meta-path indicating that there is a same attribution relationship between the author's paper and other authors' papers, and a fourth meta-path indicating that the author's paper and other authors' papers are on the same subject.

[0025] S3.2, under the same meta-path, the author nodes are aggregated using the multi-head attention mechanism to obtain the internal embedding representation of the author nodes based on the same meta-path;

[0026] S3.3, the multi-head attention mechanism is used to aggregate the author nodes under different meta-paths to obtain the long-term interest embedding representation of each author node.

[0027] In step S3.2, the author node is based on the expression of the internal embedding representation of the same meta-path:

[0028]

[0029] In the formula, Represents the author node A i In the metapath φ m The internal embedding representation below, K represents the number of heads of the multi-head attention mechanism, Represents the author node A i In the metapath φ m The set of all neighbor nodes under Represents the author node A i and author node A i′ The normalized value of the importance between i′ Represents the author node A i′ The embedding vector of .

[0030] In step S3.3, the expression of the long-term interest embedding representation of the author node is:

[0031]

[0032] In the formula, M represents the number of meta-paths, Represents the meta-path φ m The weight of Represents the meta-path φ m The set of internal embedding representations of all author nodes below.

[0033] The step S4 comprises the following steps:

[0034] S4.1, transform the work nodes and author nodes using the activation function to obtain the initial short-term interest embedding representation of each author node;

[0035] S4.2, using LSTM based on attention mechanism to capture the sequence feature representation of each author’s short-term publication history;

[0036] S4.3, the initial short-term interest embedding representation obtained in step S4.1 and the sequence feature representation obtained in step S4.2 are fused to obtain the author's short-term interest embedding representation.

[0037] In step S4.1, the expression of the initial short-term interest embedding representation is:

[0038] u c =∑ j α j d j ;

[0039] In the formula, u c represents the initial short-term interest embedding representation of the author node, α j Represents the work node P j The weight of j Represents the work node P j The corresponding embedding vector;

[0040] The work node P j The weight α j The calculation formula is:

[0041]

[0042] In the formula, v T represents the transpose of the learnable parameter matrix, u i Represents the converted author node A i The embedding representation of j Represents the converted work node P j The embedding representation of , a represents the number of author nodes.

[0043] In step S4.3, the expression of the author's short-term interest embedding representation is:

[0044] u s =W[u c ; S];

[0045] In the formula, u s represents the author’s short-term interest embedding representation, u c represents the initial short-term interest embedding representation of the author node, and S represents the sequence feature representation of the author’s recent publication history.

[0046] Beneficial effects of the present invention:

[0047] The present invention integrates the node structure, i.e., the topological relationship of each author in the knowledge heterogeneous graph, and the content information, i.e., the research interest of each author, while retaining more information, alleviating the cold start problem existing in the prediction of cooperative relationships; it integrates the long-term and short-term interests of the authors, takes into account the changes in the interests of scholars / inventors in their personal careers, and uses the long-term and short-term interest embedding representation to depict the changing trend of cooperative interests between authors, which can more accurately capture the potential cooperative interests of scholars in a certain period of time; compared with other existing academic cooperation recommendation models, it can obtain better recommendation effects and improve the accuracy of cooperative relationship prediction; the present invention can also be applied to similar technical cooperative relationship prediction tasks, and can also obtain better results. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0049] Figure 1 It is a model framework flow chart of the present invention.

[0050] Figure 2 Flowchart for obtaining long-term interest embedding representation and short-term interest embedding representation.

[0051] Figure 3 It is a partial meta-path graph of the knowledge heterogeneous graph. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] A method for predicting academic collaboration based on heterogeneous graph neural networks, such as Figure 1 and Figure 2 As shown, the following steps are included:

[0054] S1, collect the author's work information to build a work dataset, build a knowledge heterogeneous graph based on the work dataset, and divide the work dataset into a training set and a test set;

[0055] The work data set includes an academic data set and a technical data set. The academic data set is obtained by collecting data information in an academic database, and the technical data set is obtained by collecting data information in a patent database. The academic data set and the technical data set are used to construct a knowledge heterogeneous graph. The academic data set includes scholars, papers, conferences, and topics, and the technical data set includes inventors, patents, regions, and IPC classification numbers. The knowledge heterogeneous graph is represented by G, G = (V, E), V represents the node set in the knowledge heterogeneous graph, the nodes in the node set V include author nodes, work nodes, attribution nodes, and content nodes, and the set of all author nodes is represented by A, A = {A1..., A i , ..., A a}, where A i represents the i-th author node; the set of all work nodes is represented by P, where P = {P1..., P j , ..., P p}, where P j represents the jth work node; the set of all belonging nodes is represented by C, C = {C1..., C k , ..., C c}, where C k represents the kth belonging node; the set of all content nodes is represented by T, T = {T1..., T l , ..., T t}, where T l Indicates the lth content node. In the established knowledge heterogeneous graph, different nodes may correspond to the same data set or different data sets. When it is an academic data set, the author node represents the scholar, the work node represents the paper, that is, the paper name, the content node represents the paper theme, and the attribution node represents the location of the paper publication conference. When it is a technical data set, the author node represents the inventor, the work node represents the patent, that is, the patent name, the content node represents the IPC classification number, and the attribution node represents the region, that is, the national and provincial code of the patent. E represents the edge set formed between nodes in the knowledge heterogeneous graph. The edge types in the edge set E include the author-work writing relationship, the work-attribution relationship, the work-content correspondence relationship, and the work-work citation relationship.

[0056] This example is based on the academic data set provided by the Aminer database and the data information in the patent data set based on the Chinese National Patent Database. The data from 1999 to 2008 and 2009 to 2018 are selected respectively. The data is randomly divided into 10 parts according to the year, with the first 8 as training sets and the last 2 as test sets. A unique ID is assigned to each author during use to avoid confusion caused by duplicate names.

[0057] S2, using the DeepWalk algorithm and the Text-CNN model to extract the information in the knowledge heterogeneous graph established in step S1 to obtain the embedding vector of each author node, including the following steps:

[0058] S2.1, using word embedding to vectorize the information of work nodes and content nodes into corresponding text vectors as the content representation of each author's research interests;

[0059] Since the author may have published many papers or patents throughout his academic career, and different authors may have published different numbers of papers / patents, the number of papers or patents of the selected scholars is fixed to 5. If the number is greater than this, random selection is performed, and if the number is less than this, repeated selection is performed to make up for it. The existing technology uses one-hot pre-encoding and then converts it into a word vector of the corresponding dimension. Although it is simple and efficient, it ignores the author's personal research interests, and there is a cold start problem in the problem of predicting cooperative relationships.

[0060] S2.2, input the text vectors obtained in step S2.1 into the Text-CNN model for feature vector extraction, and obtain the title feature vector and text feature vector of each author;

[0061] The title feature vector is obtained by extracting the paper name or patent name, and the text feature vector is obtained by extracting the IPC classification number and the paper subject. Specifically, two parallel Text-CNN models are used as text information extractors, and the content representation, that is, the text vector corresponding to the work node and the text vector corresponding to the content node are respectively input into the two Text-CNN models. The dimension of the text vector is d and the length is n, thus forming an n*d matrix respectively. By inputting the matrix into the Text-CNN model, the title feature vector and text feature vector of each author can be obtained.

[0062] S2.3, using BiLSTM or a fully connected layer to fuse the title feature vector and the text feature vector obtained in step S2.2 to obtain the interest feature vector of each author;

[0063] The specific method is to add two feature vectors and pass them through a fully connected layer to obtain the feature representation of the author's research content.

[0064] S2.4, treating the knowledge heterogeneous graph established in step S1 as homogeneous, and using the DeepWalk algorithm to pre-train the knowledge heterogeneous graph to obtain a vector representation reflecting the topological relationship of each author node in the knowledge heterogeneous graph;

[0065] The knowledge heterogeneous graph is regarded as homogeneous, and the adjacency list information corresponding to the network structure is input into DeepWalk. The output result is the vector representation of the topological relationship between each node and other nodes. The vector representation of the author node is proposed as the structural embedding vector of the author.

[0066] S2.5, fuse the interest feature vector obtained in step S2.3 and the vector representation obtained in step S2.4 to obtain the embedding vector of each author node, so as to realize the extraction of each author's research interest, i.e., content, and the topological relationship, i.e., structure, of each author in the knowledge heterogeneous graph.

[0067] S3, using the multi-head attention mechanism to aggregate the meta-paths in the knowledge heterogeneous graph based on the author nodes to obtain the long-term interest embedding representation that characterizes the author's long-term cooperation tendency, including the following steps:

[0068] S3.1, select meta-paths with length less than N from the knowledge heterogeneous graph based on the author node;

[0069] There are many meta-paths in a heterogeneous graph, but a very long meta-path does not actually have much significance. In this embodiment, N=5, and there are four meta-paths in this application, such as Figure 3 As shown in the figure, they are APA, which indicates the cooperation relationship between the author and other authors, APPA, which indicates the citation relationship between the author and other authors, APCPA, which indicates the same affiliation relationship between the author's paper and other authors' papers, and APTPA, which indicates the same topic relationship between the author's paper and other authors' papers. Since the meta-path expresses the interactive relationship between the author and other authors in the knowledge heterogeneous graph to some extent, the meta-path is used to represent the long-term cooperation trend between authors.

[0070] S3.2, under the same meta-path, the author nodes are aggregated using the multi-head attention mechanism to obtain the internal embedding representation of the author nodes based on the same meta-path;

[0071] Under the same meta-path, different neighbor nodes may have different influences on the author node, which is reflected in the fact that an author is more inclined to cooperate with other authors who are more important. Assume that the author node pair A(i, i′) is connected through the meta-path φ m Connected, author node A i and author node A i′ The importance of The calculation formula is:

[0072]

[0073] In the formula, h i Represents the author node A obtained in step S2i The embedding vector, h i′ Represents the author node A obtained in step S2 i′ The embedding vector of i′ ∈A.

[0074] Since the author node A i and author node A i′ The importance of is asymmetric, so After that, it is normalized to obtain according to Get the author node A i About the meta-path φ m The internal embedding representation of Internal Embedding Representation The calculation formula is:

[0075]

[0076] In the formula, Represents the author node A i In the metapath φ m The set of all neighbor nodes under h i′ Represents the author node A i′ The embedding vector of , σ(·) represents the activation function. Multi-head attention is used to calculate the node-level attention, which is more effective. Specifically, the attention model is repeated n times independently. Different from the multi-head connection in GAT, the average value of the learned internal embedding representation is used as the final vector. The calculation formula is as follows, K represents the number of attention heads. In this embodiment, K=4, which means that the average value after 4 repeated experiments is the final result.

[0077] When multi-head attention is used to calculate node-level attention, the corresponding internal embedding representation is calculated as:

[0078]

[0079] S3.3, using the multi-head attention mechanism to aggregate author nodes of different meta-paths in the knowledge heterogeneous graph, and obtain the long-term interest embedding representation of each author node;

[0080] Different meta-paths have different semantic expressions, so their contributions to the potential cooperation between authors are also different. Therefore, this application uses semantic-level attention to learn the importance of different meta-paths and fuse node representations under multiple meta-paths.

[0081]

[0082] In the formula, Represents the meta-path φ m The weight of Represents the meta-path φ m The set of internal embedding representations of all author nodes obtained by step S3.2, M represents the number of meta-paths. attention(·) represents the attention value calculation operation at the meta-path level. Its purpose is to learn the importance of each edge, that is, each meta-path. This importance is recorded as Metapath φ m Importance The calculation formula is:

[0083]

[0084] In the formula, a represents the number of author nodes, q T represents the transpose of the learning parameter matrix, W represents the weight matrix, and b represents the bias vector.

[0085] After obtaining the importance of each meta-path, it is normalized by the softmax function to obtain the attention weight corresponding to each meta-path, the meta-path φ m Weight The calculation formula is:

[0086]

[0087] Where exp(·) returns the exponential of the corresponding value.

[0088] By taking the weighted sum of the weights of each meta-path obtained, we can get the long-term interest embedding representation of the author node of each meta-path. The calculation formula is:

[0089]

[0090] In the formula, u l Represents the author’s final long-term interest embedding representation.

[0091] S4, based on LSTM (Long Short-Term Memory) and attention mechanism, obtains short-term interest embedding representation that characterizes the author's short-term cooperation tendency, including the following steps:

[0092] S4.1, transform the work nodes and author nodes using the activation function to obtain the initial short-term interest embedding representation of each author node;

[0093] o j =tanh(Wd j +b); (8)

[0094] In the formula, o j Represents the work node P j The transformed embedding representation, d j Represents the work node P j The corresponding embedding vector, tanh(·) represents the activation function.

[0095] u i =tanh(Wd i +b); (9)

[0096] In the formula, u i Represents the author node A i The embedding representation of i Represents the author node A i The embedding vector of .

[0097]

[0098] In the formula, α j Represents the work node P j The weight, v T represents the transpose of the learnable parameter matrix.

[0099] u c =∑ j α j d j ; (11)

[0100] In the formula, u c Represents the initial short-term interest embedding representation of the author node. The author's recently published works are taken as input, and the size of the collection is uniformly set to 5. The embedding vector of the author node is used as the query vector, and the scoring function also selects the additive model. Finally, the author's initial short-term interest embedding representation u is obtained. c .

[0101] S4.2, using LSTM based on attention mechanism to capture the sequential feature representation of the author’s short-term publication history;

[0102] LSTM is a chain structure with repeated neural network modules. Each of its individual modules has an output of a hidden layer. The output h of this hidden layer can represent the author's interest at the current moment. LSTM takes the author's recently published works as input and outputs the author's short-term interest sequence feature representation. Since each author's current published work is often affected by the previous publication history, the attention mechanism is combined with LSTM to output each hidden state h. r With its previous implicit state h1, h2, ..., h r-1 To obtain the sequence feature representation s at different times r Finally, these features (s1, s2, ..., sr ) is fused through Convolutional Neural Networks (CNN) to obtain the sequence feature representation S of the author’s recent publication history, where r represents the time.

[0103] S4.3, concatenate the sequence feature representations obtained in step S4.1 and step S4.2 to obtain the author's short-term interest embedding representation, and the corresponding expression is;

[0104] u s =W[u c ; S]; (12)

[0105] In the formula, u s Representing the author’s short-term interest embedding.

[0106] S5, using the binary classification prediction method to build a binary classification model, according to the training set, test set, and the final long-term interest embedding representation u l And the final short-term interest embedding representation u s Obtain academic collaboration prediction models;

[0107] S5.1, concatenate the long-term interest embedding representation obtained in step S3 and the short-term interest embedding representation obtained in step S4 to obtain the author's final interest embedding representation, and the corresponding expression is:

[0108] μ=W[u l ;u s ]; (13)

[0109] Where μ represents the author’s final interest embedding representation.

[0110] S5.2, using the binary classification prediction method to build a binary classification model, the training set data and the final interest embedding representation u are input into the binary classification model for training to obtain the initial prediction model of academic cooperation;

[0111] The binary classification prediction method includes a fully connected layer, a GRU (Gated Recurrent Unit), a random forest, and a logistic regression method.

[0112] S5.3, input the test set data into the initial prediction model of academic cooperation, and use cross entropy as the loss function to optimize the initial prediction model of academic cooperation to obtain the academic cooperation prediction model.

[0113] The expression of the academic cooperation prediction model is:

[0114]

[0115] In the formula, represents the probability of collaboration between two authors predicted by the academic collaboration prediction model, μ i1 Represents the author node A i1 The final interest embedding representation, μ i2 Represents the author node A i2 The final interest embedding representation of A i2 ∈A,A i1 ∈A, DNN (Deep neural network).

[0116] Training was performed in batches on academic and technical datasets, with the dimension of input features and output embedding set to 128, the learning rate set to 0.001, Adam used as the optimizer, the batch size set to 200, the number of training iterations set to 60, the random seed set to 10, and the number of attention heads set to 4. Using F1-score and AUC (area under the curve, model evaluation index) as metrics, this application is compared with several other methods, as shown in the following table:

[0117] Table 1 Experimental results on academic datasets

[0118] Dataset Academic Dataset-1 Academic Dataset-1 Academic Dataset-2 Academic Dataset-2 Evaluation indicators AUC F1 AUC F1 RWR-CR 60.3 51.4 61.9 53.2 Metapath2vec 58.6 31.8 59.6 34.8 GraphSAGE 67.2 57.3 68.5 61.5 HetGNN 70.1 64.2 71.7 66.9 SHNE 67.2 61.2 68.3 63.9 HAN 70.2 69.7 70.1 69.7 Our model 74.3 71.5 74.1 71.4

[0119] Table 2 Experimental results on technical datasets

[0120] Dataset Technical Dataset-1 Technical Dataset-1 Technical Dataset-2 Technical Dataset-2 Evaluation indicators AUC F1 AUC F1 RWR-CR 58.1 50.2 60.2 50.2 Metapath2vec 57.2 45.7 57.6 44.1 GraphSAGE 65.9 54.2 64.1 60.3 HetGNN 69.2 61.2 67.0 63.8 SHNE 66.5 58.7 64.4 59.8 HAN 70.3 65.4 68.2 66.9 Our model 73.1 69.0 72.6 69.2

[0121] Among them, RWR-CR is based on random walk and restart algorithm, which implements random walk with meta-path; Metapath2vec (Scalable Representation Learning for Heterogeneous Networks) In this model, heterogeneous neighbor sets are formed by random walks based on meta-paths, and heterogeneous skip graphs are used to learn node representations; GraphSAGE (Graph Sample Aggregate) is a classic graph neural network model, and the feature representation of nodes is obtained by aggregating the information of adjacent nodes in a specific form (mean, pooling or LSTM); HetGNN (Heterogeneous Graph Neural Network) This model uses random walk and restart algorithms to obtain neighbor sets and learns node embedding through multi-layer neighbor aggregation; SHNE combines structural graph similarity and semantic text relevance to learn node representations in text-related heterogeneous graphs; HAN This method connects nodes of the same type through meta-paths and aggregates information from different subgraphs through attention to convert heterogeneous graphs into homogeneous networks. As can be seen from Tables 1 and 2, on academic and technical datasets, compared with the best performing HAN model and HetGNN model, this application has a 2-4% improvement in both AUC and F1-score.

[0122] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for predicting academic cooperation relationships based on heterogeneous graph neural networks, characterized in that: The steps include: S1, collect the author's work information to build a work dataset, and build a knowledge heterogeneous graph based on the work dataset; The knowledge heterogeneous graph is represented by G, G = (V, E), V represents the node set in the knowledge heterogeneous graph, E represents the edge set formed between the nodes in the knowledge heterogeneous graph, and the nodes in the node set V include author nodes, work nodes, attribution nodes and content nodes; S2, using the DeepWalk algorithm and the Text-CNN model to extract the information in the knowledge heterogeneous graph established in step S1 to obtain the embedding vector of each author node; S3, uses a multi-head attention mechanism to aggregate meta-paths in the knowledge heterogeneous graph based on author nodes to obtain the long-term interest embedding representation of each author node; S4, based on LSTM and attention mechanism, obtains short-term interest embedding representation that characterizes the author's short-term cooperation tendency; S5, using the binary classification prediction method to build a binary classification model, input the work data set into the binary classification model for training to obtain the academic cooperation prediction model; The step S3 comprises the following steps: S3.1, select meta-paths with length less than N from the knowledge heterogeneous graph based on the author node; The meta-path includes a first meta-path indicating that there is a cooperative relationship between the author and other authors, a second meta-path indicating that there is a citation relationship between the author and other authors, a third meta-path indicating that there is a same attribution relationship between the author's paper and other authors' papers, and a fourth meta-path indicating that the author's paper and other authors' papers are on the same subject. S3.2, under the same meta-path, the author nodes are aggregated using the multi-head attention mechanism to obtain the internal embedding representation of the author nodes based on the same meta-path; S3.3, using the multi-head attention mechanism to aggregate author nodes under different meta-paths to obtain the long-term interest embedding representation of each author node; In step S3.2, the author node is based on the expression of the internal embedding representation of the same meta-path: In the formula, Represents the author node A i In the metapath φ m The internal embedding representation below, k represents the number of heads of the multi-head attention mechanism, Represents the author node A i In the metapath φ m The set of all neighbor nodes under Represents the author node A i and author node A i′ The normalized value of the importance between i′ Represents the author node A i′ The embedding vector of In step S3.3, the expression of the long-term interest embedding representation of the author node is: In the formula, M represents the number of meta-paths, Represents the meta-path φ m The weight of Represents the meta-path φ m The set of internal embedding representations of all author nodes below.

2. The method for predicting academic cooperation relationships based on heterogeneous graph neural networks according to claim 1 is characterized in that: The step S2 comprises the following steps: S2.1, using word embedding to vectorize the information of each work node and content node into a corresponding text vector; S2.2, input the text vectors obtained in step S2.1 into the Text-CNN model for feature vector extraction, and obtain the title feature vector and text feature vector of each author; S2.3, using BiLSTM or a fully connected layer to fuse the title feature vector and the text feature vector obtained in step S2.2 to obtain the interest feature vector of each author; S2.4, treating the knowledge heterogeneous graph established in step S1 as homogeneous, and using the DeepWalk algorithm to pre-train the knowledge heterogeneous graph to obtain a vector representation reflecting the topological relationship of each author node in the knowledge heterogeneous graph; S2.5, fuse the interest feature vector obtained in step S2.3 and the vector representation obtained in step S2.4 to obtain the embedding vector of each author node.

3. The method for predicting academic cooperation relationships based on heterogeneous graph neural networks according to claim 1 is characterized in that: The step S4 comprises the following steps: S4.1, transform the work nodes and author nodes using the activation function to obtain the initial short-term interest embedding representation of each author node; S4.2, using LSTM based on attention mechanism to capture the sequence feature representation of each author’s short-term publication history; S4.3, the initial short-term interest embedding representation obtained in step S4.1 and the sequence feature representation obtained in step S4.2 are fused to obtain the author's short-term interest embedding representation.

4. The method for predicting academic cooperation relationships based on heterogeneous graph neural networks according to claim 3 is characterized in that: In step S4.1, the expression of the initial short-term interest embedding representation is: you c =∑ j a j d j ; In the formula, u c represents the initial short-term interest embedding representation of the author node, α j Represents the work node P j The weight of j Represents the work node P j The corresponding embedding vector; The work node P j The weight α j The calculation formula is: In the formula, v T represents the transpose of the learnable parameter matrix, u i Represents the converted author node A i The embedding representation of j Represents the converted work node P j The embedding representation of , a represents the number of author nodes.

5. The method for predicting academic cooperation relationships based on heterogeneous graph neural networks according to claim 3 is characterized in that: In step S4.3, the expression of the author's short-term interest embedding representation is: u s =W[u c ;S]; In the formula, u s represents the author’s short-term interest embedding representation, u c represents the initial short-term interest embedding representation of the author node, and S represents the sequence feature representation of the author’s recent publication history.

Images