A scientific research paper cooperation prediction method based on dynamic graph representation learning

Abstract

The present application belongs to the field of information technology, and particularly relates to a scientific research paper cooperation prediction method based on dynamic graph representation learning. The present application captures the structural information in the scientific research paper cooperation relationship network through a graph convolutional neural network to generate embedding vectors of nodes; captures edge embedding vectors representing the historical time-related information in the scientific research paper cooperation relationship network through a fully connected neural network. Then, the two kinds of vectors are fused to predict the future scientific research paper cooperation relationship, solving the problem that a part of historical information is gradually lost in the training process. Since a sequence training method such as a recurrent neural network is not used, the generation of the node and edge embedding vectors in the present application method is a static training method, the training efficiency of the model is greatly improved, and the accuracy is higher.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of information technology, and particularly relates to a scientific research paper cooperation prediction method based on dynamic graph representation learning. BACKGROUND

[0002] With the globalization of technological development and the diversification, refinement and complexity of research problems, scientific research cooperation has become one of the important ways of academic achievement output. Taking scholars as nodes of a network and the cooperation relationship between scholars as edges of the network, the scientific research cooperation behavior of scholars constitutes a typical scientific research cooperation network. Analyzing the scientific research cooperation network of scholars helps to clarify the current situation of scholar cooperation, improve the efficiency of scholar scientific research cooperation, and promote the integration of discipline resources, which is an important research topic. The scientific research cooperation network has high sparsity, and the relationship prediction thereof can realize accurate recommendation of potential collaborators in a sparse network, thereby enhancing the connection between different scholars, improving the network density, promoting the development of disciplines and the dissemination of knowledge, and ultimately effectively promoting the efficiency of scientific research cooperation.

[0003] For the future scientific research paper cooperation prediction task, it can be regarded as dynamic graph link prediction. In the current domestic and foreign literature on dynamic graph link prediction, the main method is to combine a graph neural network model and a recurrent neural network model, on the one hand, the graph neural network model is used to capture the graph structure information of each time period subgraph, and on the other hand, the recurrent neural network model is used to capture the time-related new information of the graph, such as EloveGCN. This method uses a recurrent neural network to use the historical information of the previous subgraph for training of the next subgraph embedding vector, which has two problems, first, a part of the historical information is gradually lost in the above training process; second, the training efficiency is low due to the time sequence information transmission process. SUMMARY

[0004] The present application aims at the deficiencies in the prior art, and provides a scientific research paper cooperation prediction method based on dynamic graph representation learning, which solves the problem of gradual loss of a part of historical information in the training process and improves the training efficiency.

[0005] To achieve the above application purposes, the application adopts the following technical solutions:

[0006] A scientific research paper cooperation prediction method based on dynamic graph representation learning comprises the following steps

[0007] S1, data is acquired and processed to obtain an initial adjacency matrix, an initial feature vector and an initial scientific research paper cooperation history information vector; the acquired data comprises a scientific research paper cooperation network original adjacency matrix, an original feature vector and a scientific research paper cooperation history information vector;

[0008] S2, inputting the initial adjacency matrix A T and the initial feature vector X T into the graph convolutional neural network to obtain a node embedding vector; inputting the initial scientific research paper collaboration history information vector into the fully connected neural network to obtain an edge embedding vector through the fully connected neural network;

[0009] S3, constructing and training a scientific research paper collaboration relationship prediction model; inputting the node embedding vector and the edge embedding vector obtained in S2 into the trained scientific research paper collaboration relationship prediction model to predict future scientific research paper collaboration.

[0010] Further, the S1 includes the following sub-steps

[0011] S1.1, obtaining data:

[0012] taking the authors of scientific research paper collaboration as nodes and taking the scientific research paper collaboration relationship as an undirected edge, constructing a scientific research paper collaboration network through subgraphs of each time period wherein t is the number of time snapshots, G t represents a subgraph of the t time period, and ε={E1, E2,..., E t} represents the edge set of the group of continuous graph snapshots, represents the node set of the group of continuous snapshots.

[0013] adopting to represent the adjacency matrix of a group of continuous graph snapshots, and the node attribute matrix corresponding to the adjacency matrix is For the adjacency matrix, if there is cooperation between scientific researchers in a certain time period, there is a connection between the corresponding nodes, and the corresponding position of the adjacency matrix is 1, otherwise it is 0; for the node attribute matrix, a one-hot encoding is used to artificially set the attribute vector of each node;

[0014] S1.2, data preprocessing:

[0015] processing the original adjacency matrix set and the corresponding attribute matrix set in S1.1 to obtain the initial adjacency matrix A T and the initial feature vector X T .

[0016] The edge set of each subgraph snapshot in S1.1 represents the existence of cooperation relationship in each time period, and an initial scientific research paper collaboration history information vector S 0 is generated.

[0017] Further, the method for generating the initial scientific research paper collaboration history information vector in S1.2 is:

[0018] S1.2.1 integrate the set of edges that exist in each time period,

[0019] E T = E1∪E2∪…∪E t-1

[0020] S1.2.2 for any edge e ab ∈E T time series vector S ab = [s1 s2…s t-1 ], initialized as follows:

[0021]

[0022] S 0 = {S e |e∈E T}.

[0023] Further, the S3 trains the scientific paper collaboration relationship prediction model according to the following steps:

[0024] S3.1, 1:1 positive and negative sampling is performed on the scientific paper collaboration in the last time period, the positive sampling refers to the result set E pos , and the negative sampling refers to the result set E neg that does not exist;

[0025] A scientific paper collaboration relationship prediction model is established, which is represented by the following formula:

[0026]

[0027] wherein e ij represents an edge, E T represents a historical edge set, S ij is the trained embedding vector corresponding to the edge e ij in S, z i and z j represent the embedding vectors of node i and node j in the trained node embedding vector matrix Z, represents a weight control parameter for controlling the embedding vector information of the node and the time-related information provided by the embedding vector of the edge, λ∈[0,1], represents the predicted existence probability of the edge e ij in the future time period, sigmoid(*) represents a sigmoid function, and sum(*) represents a sum function;

[0028] S3.2, according to the node embedding vector and the edge embedding vector obtained by S2 and the scientific paper collaboration relationship prediction model established by S3.1, the scientific paper collaboration probability in the last time period is predicted;

[0029] S3.3, adopt cross-entropy loss function as loss function, introduce the positive and negative sampling result set obtained in S3.2 into the cross-entropy loss function, calculate the loss of the scientific research paper cooperation relationship prediction model, and the specific calculation is as follows:

[0030]

[0031] Wherein, p e represents whether the edge e exists, represents the probability of predicting the existence of edge e;

[0032] S3.4, taking the minimum loss function as the goal, using gradient descent method and Adam optimizer to train the scientific research paper cooperation relationship prediction model;

[0033] S3.5, judge whether the loss function value converges, if yes, the training is completed, otherwise return to step S3.4.

[0034] Further, the S3 utilizes the trained scientific research paper cooperation relationship prediction model to predict the future scientific research paper cooperation situation, and the specific operation is as follows:

[0035] Fusing the trained node vector and edge embedding vector obtained in S2, the existence probability of future scientific cooperation relationship is obtained

[0036] Judge the existence probability of the research paper cooperation relationship Whether greater than 0.5, if yes, it is determined that the scientific researcher i and the scientific researcher j will cooperate in the future, if not, there is no cooperation between them in the future.

[0037] Further, the graph convolutional network of the S2 is a two-layer graph convolutional neural network, and the formula of each layer is as follows:

[0038] H (k+1) =σ(H k ,A T |W k )=σ(D -1 / 2 A T D 1 / 2 H k W k )

[0039] Wherein, sigma represents the activation function, and LeakyReLU activation function is selected, the input of the first layer of graph convolutional layer is the initial attribute matrix X T =H 0 ; The merged graph adjacency matrix information A T and the attribute information X TAfter inputting the defined graph convolution layer model, the output of the second layer graph convolution layer can be obtained, and the output is a node embedding vector matrix, that is, Z = H (2) = {z1, z2, …, z N} T .

[0040] Further, the full connection neural network of S2 is also a two-layer full connection neural network, and each layer is formulated as follows:

[0041]

[0042] y (k+1) = σ (W (k+1) y (k) + b (k+1) )

[0043] Wherein, sigma represents an activation function, the linear rectifier unit ReLU is selected as the activation function in the method, y (k) represents the output result of the kth full connection layer, the input of the first full connection layer is the initial cooperation history information vector matrix S 0 =y 0 . The cooperation history information embedding vector output by the second full connection layer, that is, the last full connection layer, is S = y 2 .

[0044] The application captures the structural information in the scientific paper cooperation relationship network to generate node embedding vectors through the graph convolution neural network; the edge embedding vectors representing the historical time related information in the scientific paper cooperation relationship network are captured through the full connection neural network. Then, the two kinds of vectors are fused to predict the future scientific paper cooperation relationship, and the problem that part of the historical information is gradually lost in the training process is solved. Since the sequence training method of the recurrent neural network is not used, the generation of the node and edge embedding vectors in the method is a static training method, the training efficiency of the model is greatly improved, and the accuracy is higher.

[0045] Compared with the prior art, the prediction method of the application can accurately and efficiently predict the future scientific paper cooperation relationship, realize the accurate recommendation of potential collaborators in the sparse network, enhance the connection between different scholars, improve the network density, promote the development of disciplines and knowledge dissemination, and finally effectively promote the efficiency of scientific cooperation. BRIEF DESCRIPTION OF DRAWINGS

[0046] Figure 1 A scientific paper cooperation prediction method flow chart based on dynamic graph representation learning is proposed in the application;

[0047] Figure 2 A method work flow framework diagram is proposed in the application;

[0048] Figure 3 Fig. 1 is a schematic diagram of a scientific paper collaboration network instance in an embodiment of the present application;

[0049] Figure 4 Fig. 4 is a sensitivity analysis result graph of node, edge embedding vector weight setting in the present application;

[0050] Figure 5 Fig. 5 is a comparative experiment result schematic diagram after adjusting the length of the training set in the present application. DETAILED DESCRIPTION

[0051] The specific embodiments of the present application are described below to facilitate those skilled in the art to understand the present application, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, as long as various changes are within the spirit and scope of the present application defined and determined by the appended claims, all applications utilizing the concept of the present application are within the scope of protection.

[0052] The embodiments of the present application are described in detail below with reference to the accompanying drawings.

[0053] Figure 2 Fig. 1 is a schematic diagram of a scientific paper collaboration network instance in an embodiment of the present application;

[0054] As shown in the figure, the scientific paper collaboration prediction method based on dynamic graph representation learning provided by the embodiment includes the following steps: Figure 1

[0055] S1, obtaining an initial adjacency matrix, a feature vector and a history information vector of scientific paper collaboration;

[0056] The step S1 includes the following sub-steps:

[0057] S1.1, taking the authors of scientific paper collaboration as nodes and the scientific paper collaboration relationship as undirected edges, constructing a scientific paper collaboration network through subgraphs of each time period wherein t is the number of time snapshots, G t represents a subgraph of t time period, ε={E1, E2, …, E t} represents the edge set of this group of continuous graph snapshots,​ A set of nodes representing this set of consecutive snapshots.

[0058] Adopting An adjacency matrix representing a set of consecutive snapshots of graphs, the first-order adjacency matrix corresponding to a node attribute matrix For the adjacency matrix, if there is cooperation between researchers in a certain period of time, there is a connection between the corresponding nodes, the adjacency matrix corresponding position is 1, otherwise 0. For the node attribute matrix, the nodes of the current scientific paper cooperation network usually have no specific attributes, and the embodiment artificially sets the attribute vector for each node by using one-hot encoding.

[0059] S1.2, using the original adjacency matrix set in S1.1 And the attribute matrix set Respectively remove the last time period matrix to generate the initial adjacency matrix A T And the feature vector X T The synthesis method is:

[0060] A T =A1+A2+…+A t-1

[0061]

[0062] S1.3, using the edge set ε of each subgraph snapshot in S1.1 to represent the existence of cooperation relationship in each time period, and generating an initial scientific paper cooperation history information vector. Specifically:

[0063] Integrate the set of edges that exist in each time period:

[0064] E T =E1∪E2∪…∪E t-1

[0065] According to the integration result, for any edge e ab ∈E T Time series vector S ab =[s1 s2…s t-1 ] is initialized as follows:

[0066]

[0067] S 0 ={S e |e∈E T}

[0068] The obtained S 0 is the initial scientific paper cooperation history information vector matrix.

[0069] In this embodiment, the scientific paper collaboration dataset used is publicly available, and this academic collaboration dataset contains 315 researchers' scientific paper collaboration from 2000 to 2009, and each year is divided into a subgraph as a time period, and each subgraph node represents an author and an edge corresponds to a scientific paper collaboration relationship. The specific information of the dataset is shown in Table 1.

[0070] Table 1 Statistics of scientific collaboration network data

[0071]

[0072] The above table is the specific information of the CLOAB dataset, and the length of the training set in this example is 7. We select the network of the first 7 time periods for training to predict the scientific paper collaboration in the 8th time period.

[0073] For ease of understanding, the present embodiment is described as follows. Figure 3 A training set containing 5 time periods and 5 user scientific paper collaboration network instances is constructed. The attribute matrix is generated by one-hot encoding.

[0074] The adjacency matrix and attribute matrix of the network of the first time period are shown in Tables 2 and 3.

[0075] Table 2 Adjacency matrix of the network of the first time period

[0076]

[0077] Table 3 Node attribute matrix of the first time period

[0078]

[0079] The adjacency matrix and attribute matrix of the network of the second time period are shown in Tables 4 and 5.

[0080] Table 4 Adjacency matrix of the network of the second time period

[0081]

[0082] Table 5 Node attribute matrix of the second time period

[0083]

[0084] The adjacency matrix and attribute matrix of the network of the third time period are shown in Tables 6 and 7.

[0085] Table 6 Adjacency matrix of the network of the third time period

[0086]

[0087] Table 7 Node attribute matrix of the third time period

[0088]

[0089] The adjacency matrix and attribute matrix of the network in the fourth time period are shown in Tables 8 and 9.

[0090] Table 8 Adjacency matrix of the network in the fourth time period

[0091]

[0092] Table 9 Node attribute matrix in the fourth time period

[0093]

[0094] The adjacency matrix and attribute matrix of the network in the fifth time period are shown in Tables 10 and 11.

[0095] Table 10 Adjacency matrix of the network in the fifth time period

[0096]

[0097] Table 11 Node attribute matrix in the fifth time period

[0098]

[0099] In the training, the subgraphs of the time periods except the last time period of the training set are selected to form the initial adjacency matrix, the initial feature matrix, and the initial historical vector matrix of the scientific paper cooperation relationship in step S1.

[0100] The initial adjacency matrix and the feature matrix after the synthesis of the last time period are shown in Tables 12 and 13.

[0101] Table 12 Adjacency matrix after synthesis

[0102]

[0103] Table 13 Node attribute matrix after synthesis

[0104]

[0105] The initial historical matrix of the scientific paper cooperation relationship formed by removing the last time period is shown in Table 14.

[0106] Table 14 Initial cooperation relationship historical matrix

[0107]

[0108] S2, using the data processed in S1 to train the model parameters, to obtain the embedding vectors of the nodes and edges that capture the spatiotemporal information of the scientific paper cooperation network. The specific operation of step S2 is as follows:

[0109] The initial adjacency matrix A T and the initial feature vector X T The initial scientific paper collaboration history information vector is input into a full connection neural network to obtain an edge embedding vector.

[0110] The graph convolutional neural network has the functions of smoothing and propagating information. The graph convolutional neural network is a two-layer graph convolutional neural network, and the formula of each layer is as follows:

[0111] H (k+1) =σ(H k ,A T |W k )=σ(D -1 / 2 A T D 1 / 2 H k W k )

[0112] In this example, the LeakyReLU activation function is selected for σ, and the input of the first layer is Since two-layer graph convolutional neural networks are used, two corresponding parameter matrices The final obtained node embedding vector matrix is

[0113] The two-layer full connection neural network is used to train the edge time sequence information vector, and the formula of each layer is as follows:

[0114] y (k) =σ(W (k) y (k-1) +b (k) )

[0115] y (k+1) =σ(W (k+1) y (k) +b (k+1) )

[0116] In this example, the ReLU activation function is selected for σ, and the input of the two-layer full connection network is 564 is the number of collaboration relationships that appear in the training set except for the last time period, and 6 represents the number of time periods. The parameter matrices corresponding to the two-layer full connection layers are The two corresponding bias matrices are The collaboration history information embedding vector output by the second layer, that is, the last full connection layer, is

[0117] S3, fuse the node embedding vector and the edge embedding vector obtained in S2 to predict the future scientific paper collaboration. The detailed process is as follows:

[0118] Construct and train the scientific paper collaboration relationship prediction model:

[0119] S3.1, establish a scientific paper collaboration relationship prediction model, which is expressed by the following formula:

[0120]

[0121] E T is the set of edges that exist in each time period, || represents vector splicing, and are the embedding vectors of the nodes corresponding to the edge (i, j), is the trained historical collaboration information corresponding to the edge (i, j).

[0122] As can be seen from the above formula, the prediction method used in this embodiment needs to first determine whether the edge e ij exists in the historical edge set E T If not, it is calculated by the node similarity method, otherwise we first multiply the embedding vectors of the nodes, complete the first step of calculating the edge existence probability, at this time it is equivalent to obtaining a similarity vector between the node embedding vectors. Then, the trained edge existence sequence embedding vector is spliced with the product of the two node embedding vectors obtained at present, which is equivalent to expanding the edge features, and then the edge existence probability is calculated. λ represents a weight control parameter that controls the information of the node embedding vector and the time-related information provided by the edge embedding vector, λ ∈ [0, 1], represents the existence probability of the predicted edge e ij , sigmoid(*) represents the sigmoid function, and sum(*) represents the sum function.

[0123] The scientific paper collaboration in the last time period is 1:1 positive and negative sampling for loss function calculation. The result set of sampling positive edges, i.e. existing edges, is The result set of sampling negative edges, i.e. non-existing edges, is

[0124] S3.2, according to the node embedding vectors and edge embedding vectors obtained in S2, and the scientific paper collaboration relationship prediction model established in S3.1, predict the scientific paper collaboration probability in the last time period;

[0125] S3.3, construct a loss function. This method uses a cross-entropy loss function as the loss function, and introduces the positive and negative sampling result sets obtained in S3.2 into the cross-entropy loss function to calculate the loss of the scientific paper collaboration relationship prediction model. The specific calculation is as follows:

[0126]

[0127] where p erepresenting whether edge e exists, representing the probability that edge e exists.

[0128] S3.4, using gradient descent and Adam optimizer to train the scientific paper collaboration relationship prediction model with the minimum loss function as the goal;

[0129] S3.5, judge whether the loss function value converges, if yes, the training is completed, otherwise return to step S3.4.

[0130] Fusion of the two trained embedding vectors obtained in S2 to predict future scientific paper collaboration:

[0131] Using the trained scientific paper collaboration relationship prediction model to fuse the node vector and edge embedding vector obtained in S2 to obtain the existence probability of future scientific collaboration The calculation formula is as follows:

[0132]

[0133] In this example, λ = 0.3, E T is the set of edges that exist in each time period, || represents vector splicing, Both are the embedding vectors of the nodes corresponding to edge (i, j), is the trained historical collaboration information corresponding to edge (i, j), sigmoid(*) represents the sigmoid function, and sum(*) represents the sum function. That is, the existence probability of edge (i, j) in the test set, that is, the future time period.

[0134] Judge whether the existence probability of scientific paper collaboration relationship obtained in S3.1 is greater than 0.5, if yes, it is determined that scientific researcher i and scientific researcher j will have cooperation in the future, if not, there will be no cooperation between them in the future.

[0135] The current mainstream dynamic graph link prediction method is used as the baseline method, including the dynamic graph application method of classic graph embedding method GCN, GAT and GCN-GRU, GC-LSTM, EloveGCN, HTGN specially used for dynamic graph, among which EloveGCN has two versions of -H and -O, the former uses GRU for recurrent neural network, and the latter uses LSTM. The indicators for measuring link prediction task are AUC and MAP, and the experimental results are shown in Table 15.

[0136] Table 15 Link prediction result table

[0137] ​

[0138] According to Table 15, we can know that the static graph method GCN, GAT is much worse than other methods specially designed for dynamic graphs, and our method is optimal in AUC and MAP, and has better prediction ability.

[0139] In addition, the current baseline method all uses recurrent neural network as the way to capture time information, and the training efficiency of recurrent neural network is low. Our method gives up this idea and puts the time-related information in the edge embedding vector, adopts the static training method, and has higher training speed. Table 16 is the average time required for training 100 epochs of each model, which is in ms. Here, only the methods specially designed for dynamic graphs are calculated, and the GCN and GAT static graph methods are not calculated.

[0140] Table 16 Average time required for one round of training

[0141]

[0142] From Table 16, we can know that our method greatly reduces the training time of the model and greatly improves the training efficiency.

[0143] Figure 4 It shows the influence of the weight setting of the node and edge of the present application on the test result when predicting the future cooperation relationship, that is, the sensitivity of the method to λ. The larger λ is, the greater the influence of the node embedding vector on the prediction result, and the smaller the influence of the edge embedding vector on the prediction result. Through Figure 4 We can find that when λ = 0.4, the prediction effect in the present example is best. In addition, when λ = 0, it means that the edge embedding vector is used for prediction, and when λ = 1, it means that the node embedding vector is used for prediction. The prediction effect of the two cases is obviously lower than that of the other cases, which shows that the two together will have a positive impact. Therefore, the Figure 4 .

[0144] Figure 5 It shows the change of the prediction effect of the present application and two suboptimal baseline methods HTGN and EloveGCN under the condition of different training set lengths, wherein the results of EloveGCN are the optimal values of -H and -O versions. It can be seen that with the decrease of the training set length, the prediction accuracy of the three methods decreases, which shows that rich historical information has a positive impact on predicting future scientific cooperation. It can also be seen that the present method is basically optimal at each length, which shows that it has good robustness. Therefore, the Figure 5 .

[0145] It can be seen from the above that the scientific paper cooperation prediction method based on dynamic graph representation learning has the advantages of high prediction accuracy and fast training speed. In actual production environment, given the existing scientific paper cooperation network historical information, the future scientific paper cooperation relationship can be predicted by using the method, the relationship prediction of the scientific paper cooperation network can realize the accurate recommendation of potential collaborators in the sparse network, thereby enhancing the connection between different scholars, improving the network density, promoting the development of disciplines and knowledge dissemination, and finally effectively promoting the efficiency of scientific cooperation.

Claims

1. A method for predicting scientific paper collaboration based on dynamic graph representation learning, characterized in that, The method comprises the following steps: S1, obtaining data and processing the data to obtain an initial adjacency matrix, an initial eigenvector and an initial history information vector of scientific paper cooperation; the obtained data comprises an original adjacency matrix of a scientific paper cooperation network, an original eigenvector and a history information vector of scientific paper cooperation; S2, inputting the initial adjacency matrix and the initial eigenvector into a graph convolution neural network to obtain a node embedding vector through the graph convolution neural network; inputting the initial history information vector of scientific paper cooperation into a fully connected neural network to obtain an edge embedding vector through the fully connected neural network; S3, constructing and training a scientific paper cooperation relationship prediction model; inputting the node embedding vector and the edge embedding vector obtained in S2 into the trained scientific paper cooperation relationship prediction model to predict future scientific paper cooperation; the training method of the scientific paper cooperation relationship prediction model is as follows: S3.1, 1:1 positive and negative sampling is performed on the cooperation of scientific research papers in the last time period, and the positive sampling refers to the result set E with edges pos , and the negative sampling refers to the result set E without edges neg ; The scientific paper cooperation relationship prediction model is established, and the prediction model is expressed by the following formula: where e ij denotes an edge, E T denotes a historical edge set, S ij is an edge e ij corresponding trained embedding vector, z i and z j denotes the embedding vector of node i and node j in the trained node embedding vector matrix Z, denotes the weight control parameter of the control node embedding vector information and the time-related information provided by the edge embedding vector, λ ∈ [0, 1], denotes the predicted existence probability of edge e ij in the future time period, sigmoid(*) denotes the sigmoid function, and sum(*) denotes the sum function; S3.2, predicting the scientific paper cooperation probability of the last time period according to the node embedding vector and the edge embedding vector obtained in S2 and the scientific paper cooperation relationship prediction model established in S3.1; S3.3, using a cross-entropy loss function as a loss function, introducing the positive and negative sampling result set obtained in S3.2 into the cross-entropy loss function, and calculating the loss value of the scientific paper cooperation relationship prediction model, and the specific calculation is as follows: where p e represents whether edge e exists, represents the probability that edge e exists; S3.4, taking the minimum loss function as the target, using the gradient descent method and the Adam optimizer to train the scientific paper cooperation relationship prediction model; S3.5, judging whether the loss function value converges or not, if yes, the training is completed, otherwise returning to step S3.

4.

2. The scientific paper collaboration prediction method based on dynamic graph representation learning according to claim 1, characterized in that, The S1 comprises the following sub-steps: S1.1, obtaining data: The scientific paper cooperation network is constructed by taking the authors of scientific papers as nodes and taking the cooperation relationship of scientific papers as undirected edges through subgraphs of each period in the past where t is the number of time snapshots, G t denotes the subgraph of the t period, and ε = {E1, E2,..., E t} denotes the edge set of the group of continuous graph snapshots, denotes the node set of the group of continuous snapshots; Adopting The adjacent matrix corresponding to the node attribute matrix is For the adjacent matrix, if there is cooperation between scientific researchers in a certain period of time, there is a connection between the corresponding nodes, and the corresponding position of the adjacent matrix is 1, otherwise 0; for the node attribute matrix, the attribute vector of each node is artificially set by one-hot encoding; S1.2, data preprocessing: The original adjacency matrix set in S1.1 and its corresponding attribute matrix set is processed to obtain an initial adjacency matrix A T and an initial eigenvector X T ; The edge set ε of each subgraph snapshot in S1.1 represents the existence of the cooperative relationship in each time period, and is used to generate the initial scientific paper cooperation history information vector S 0 .

3. The scientific paper collaboration prediction method based on dynamic graph representation learning according to claim 2, characterized in that, The initial history information vector of scientific paper cooperation in S1.2 is generated by the following method: Integrate the edge set existing in each time period: E T = E1 U E2 U... U En t-1 For any edge e ab ∈E T Time series vector S ab = [s1 s2 … s t-1 ] is initialized as follows:

4. The scientific paper collaboration prediction method based on dynamic graph representation learning according to claim 1, characterized in that, The specific operation of the trained scientific paper cooperation relationship prediction model in S3 for predicting future scientific paper cooperation is as follows: S2. obtaining the existence probability of the future scientific research cooperation relationship by fusing the trained node vector and the edge embedding vector obtained in S1. determining the probability of existence of a scientific research collaboration whether it is greater than 0.5, if so, determining that scientific researcher i and scientific researcher j will have a collaboration in the future, if not, they will not have a collaboration in the future.

5. The scientific paper collaboration prediction method based on dynamic graph representation learning according to claim 1, characterized in that, The graph convolution neural network in S2 is a two-layer graph convolution neural network, and the formula of each layer is as follows: H (k+1) = σ(H k ,A T |W k ) = σ(D -1 / 2 A T D 1 / 2 H k W k ) Wherein, sigma represents the activation function, and LeakyReLU activation function is selected, and the input of the first layer graph convolution layer is the initial attribute matrix X T = H 0 ; the adjacency matrix information A T of the merged graph and the attribute information X T of the corresponding node are input into the defined graph convolution layer model, and the output of the second layer graph convolution layer can be obtained, and the output is the embedding vector matrix of the node, that is, Z = H (2) = {z1, z2,..., z N} T .

6. The scientific paper collaboration prediction method based on dynamic graph representation learning according to claim 1, characterized in that, The fully connected neural network in S2 is also a two-layer fully connected neural network, and the formula of each layer is as follows: y (k) = σ(W (k) y (k-1) + b (k) ) y (k+1) = σ(W (k+1) y (k) + b (k+1) ) Wherein, sigma represents an activation function, the method selects a linear correction unit ReLU as the activation function, y (k) represents the output result of the kth fully connected layer, the input of the first fully connected layer is the initial cooperation history information vector matrix S 0 = y 0 , and the cooperation history information embedding vector output by the second, that is, the last fully connected layer is S = y 2 .

Images