A method and system for constructing user portraits based on a scientific research network

Abstract

The present invention discloses a method and system for constructing a user portrait based on a scientific research network. A scientific research network diagram is constructed based on co-authorship data of papers. Input the feature vectors of a group of nodes, calculate the attention coefficients between nodes through a layer of multi-head attention mechanism to construct a probability transition matrix, use a layer of single-head attention to calculate the updated features of nodes, predict the node labels through a classification function, and calculate the loss of feature propagation. Input the scientific research network diagram and some known node label vectors, and use the label propagation algorithm to calculate the loss of label propagation. The custom loss function is obtained by summing the losses of feature propagation and label propagation, and the label prediction result is corrected, and finally the predicted scholar research field labels are output. The invention solves the problem of the single way of obtaining labels for the user portrait, makes full use of the feature information, label information of scholars and the associated structure information between scholars, complements the label information of unknown nodes, and improves the accuracy of label prediction.

Description

Technical Field

[0001] The present invention relates to a method and system for constructing a user portrait based on a scientific research network, belonging to the technical field of data processing. Background Art

[0002] With the development of the scientific research field and the rapid expansion of the Internet, a large number of scholars have joined the scientific research network, and the academic resources in the scientific research social network have shown an explosive growth, resulting in the problem of "information overload". Information overload restricts the effective information acquisition ability of scientific research personnel. How to find relevant papers and scholar information in the vast amount of academic resources has become an urgent problem to be solved in the scientific research social network.

[0003] User portraits provide an ideal solution to the "information overload" problem. At present, the mainstream methods for constructing user portraits for social networks mainly include the propagation method based on knowledge graphs, the propagation method based on graph neural networks, and the label propagation algorithm based on natural language processing. However, the online social interactions of ordinary users can construct a social network graph through data such as forwarding, commenting, liking, and browsing between users. However, the social interactions between scientific research personnel can only establish connection relationships through citation and co-authorship data of papers, and the method is relatively single, resulting in sparse edge density and weak correlation between scholars. This makes the method for constructing user portraits for social networks unable to fully mine the correlation structure information of scholars and reduces the accuracy of label prediction. Summary of the Invention

[0004] Object of the Invention: Aiming at the problem that the method for constructing a user portrait for a social network cannot fully utilize the correlation structure information between users, the present invention provides a method and system for constructing a user portrait based on a scientific research network, considering using a label propagation algorithm to improve the graph attention network. Since the label propagation algorithm assumes that users have the same labels, it can increase the degree of association between users and effectively solve the problem of sparse edge density in the scientific research social network. Taking the scientific research social network as the starting point, the correlation structure, features, and label information of adjacent users are fused to predict the labels of unknown users. The method of the present invention breaks the limitation of the single way of obtaining labels for user portraits, fully mines the correlation between scientific research personnel, and improves the user portrait from the social dimension of scholars.

[0005] Technical Solution: A method for constructing a user portrait based on a scientific research network includes the following steps:

[0006] (1) Data preprocessing: Based on the co-authorship data information of papers, taking scholars as nodes and the co-authorship relationship between scholars as edges, construct a scientific research social network graph, process the node features into feature vectors through one-hot encoding, and convert the labels of known nodes into label vectors through one-hot encoding.

[0007] (2) Feature Propagation: The structure of the scientific research social network graph and node features are used as the input of the feature propagation layer. The feature propagation layer embeds a hierarchical attention mechanism in the graph convolutional neural network. The specific operation is to calculate the attention similarity coefficients for adjacent paired nodes using a layer of multi-head attention mechanism. After normalizing the attention similarity coefficients, a probability transition matrix is obtained. The probability transition matrix is linearly transformed with the original feature vector of the scholar to obtain the updated scholar features. Finally, a layer of single-head attention is used to fuse the updated scholar features, and the softmax activation function is used for label classification to predict the label of the active research field of the node.

[0008] (3) Label Propagation: The structure of the scientific research social network graph and the label vectors of some known scholars are used as the input of the label propagation layer. The label propagation layer refers to the label propagation algorithm, and the propagation of labels is achieved through the label propagation algorithm.

[0009] (4) Label Prediction of User Portrait: The sum of the losses calculated by the feature propagation layer and the label propagation layer is used to obtain a custom loss function. Minimizing it helps the feature propagation layer learn the optimizer, and the optimizer feedback is sent to the feature propagation layer to learn the final node representation. The final output result is the predicted research field label of the scholar.

[0010] During the label propagation process, the labels of the labeled data are kept unchanged, and the labels are passed to the unlabeled data through the propagation matrix. Finally, the label propagation of the nodes tends to be stable, and at this time, the loss of label propagation is calculated through the cross-entropy loss function.

[0011] In the data preprocessing, based on the information of co-authored paper data, scholars are used as nodes, and the co-authored paper relationship between scholars is used as edges to construct a scientific research social network graph that can quantify the cooperation relationship of scientific research users.

[0012] The scientific research social network graph is represented as G=(V, A, X, L), where G represents the name of the scientific research social network graph, V represents the node set, representing scholars; A represents the adjacency matrix, representing the co-authored relationship between scholars; X represents the feature set of scholars, that is, the node feature set, and the node features represent the keywords of each author's papers. The construction method of the feature vector is: using one-hot encoding to mark the included features as 1 and the non-included features as 0; L represents the label set owned by some scholars, and the labels represent the most active research field labels of scholars. The construction method of the label vector is: using one-hot encoding to mark the appeared labels as 1 and the non-appeared labels as 0.

[0013] During the feature propagation process, a hierarchical attention mechanism is added to the structure of the graph convolutional neural network (GCN). First, the multi-head attention mechanism layer is used to aggregate the features of neighbor nodes, realizing the adaptive matching of weights according to different cooperation relationships, obtaining attention coefficients, and obtaining the probability transition matrix through a normalization operation. Then, the original feature vectors of each node are multiplied by the probability transition matrix to calculate the updated feature information of each node. Finally, a single-head attention mechanism is used to fuse the updated node features obtained, and the softmax activation function is used for label classification to predict the research field labels of node activities. The feature propagation further includes the following steps:

[0014] 2.1) Input: The scientific research social network graph and a set of node feature vectors are used as the input of the feature propagation layer. The set of feature vectors of scholars is where N is the number of nodes and F is the dimension of node features;

[0015] 2.2) Calculate the probability transition matrix: Taking a certain node as the center, calculate the attention similarity coefficient of its adjacent nodes. Then this node is called the central node, and the adjacent nodes are called neighbor nodes. The multi-head attention mechanism is used to calculate the attention similarity coefficient between each central node and its first-order adjacent nodes, and through the softmax activation function for normalization operation, the probability transition matrix μ ij is obtained. The calculation formula is as follows:

[0016]

[0017] where the subscripts i and j represent nodes i and j, represents the feature vector of node i, W represents a trainable weight matrix, || represents the concatenation operation, represents the learnable attention parameter; represents the linear transformation of the input feature vector. Through the || operation, the transformed feature vectors of nodes i and j are concatenated together, and a dot product operation is performed with the attention weights to calculate the attention similarity coefficient between nodes i and j. This coefficient represents the importance of node j to node i. Finally, the softmax function is used to normalize the attention similarity coefficient to obtain the probability transition matrix;

[0018] 2.3) Feature propagation: Since there are many node features in the scientific research network social graph, in order to improve the calculation efficiency, the multi-head attention mechanism is used for feature propagation. The essence of multi-head attention is to concurrently execute the calculation of probability weights between neighbors. The feature vector of the node is combined with the probability transition matrix calculated through the multi-head attention mechanism, that is, the updated feature of the central node is the weighted average of the features of adjacent nodes, represents the updated feature of node i, and the calculation formula is:

[0019]

[0020] Among them represents the probability transition matrix obtained through the k-th head of attention calculation. ij, j ∈ N(i) represent node i and its neighbor nodes represents the feature of neighbor node j, and σ(·) represents the activation function; the formula means that first, the partial updated feature of node i is obtained through the multi-head attention mechanism, and then the partial updated features obtained by the k-head attention mechanism are concatenated to obtain the updated feature of node i

[0021] 2.4) Fuse the updated features of each node through a single-head attention network layer, output a set of updated feature vectors, and output updated features according to the input N node features. Let the node feature dimension of this new predicted feature vector be F’, which is expressed as

[0022] 2.5) Label prediction: Use the softmax activation function to predict the research field labels corresponding to the node features, use the cross-entropy function to calculate the error loss between the predicted labels and the true labels, and use the loss function and the gradient descent function to optimize the feature propagation process to achieve the scholar research field label prediction task

[0023] During the label propagation process, the basic assumption of the label propagation algorithm (LPA) is that connected nodes may have the same label, that is, connected scholars may have the same research direction or research interest. Therefore, iteratively propagating labels along the edges can supplement the missing label information of nodes, and further include the following content

[0024] 3.1) Model input: Take the label vector L and the adjacency matrix A in the scientific research social network graph G as the input of the label propagation process. At initialization, the labeled nodes are set as one-hot vectors, and the unlabeled nodes are set as zero vectors

[0025] 3.2) Label propagation: Let be the label matrix in the k-th iteration, where iteration refers to continuously repeating the label propagation process, and the element represents the predicted distribution of the node labels in the k-th iteration. The calculation formula for label propagation after k iterations is as follows

[0026] L k+1 = D -1 AL (k)

[0027] Among them, A represents the adjacency matrix, D is a diagonal matrix, and the value is obtained by summing the corresponding row elements in the adjacency matrix A. The formula means that labels are passed along the edges

[0028] 3.3) Retain the labeled nodes: Since the label information of some known scholars is determined, after the propagation is completed, the labels of the known user nodes need to be reset to the original labels;

[0029] 3.4) The final label in LPA is the weighted average of neighbor nodes, and the cross-entropy loss function in tensorflow is used to calculate the loss of label propagation.

[0030] In the label prediction of the user profile, the loss functions of feature propagation and label propagation are summed to construct a custom loss function, and the loss function is used to optimize the label prediction result of feature propagation to improve the accuracy of label prediction. It further includes the following content;

[0031] Extract the features, labels, and associated structure information of adjacent nodes through feature propagation and label propagation, and perform label prediction on the research fields of scholars. The calculation formula of the custom loss function loss is as follows:

[0032]

[0033] Among them, represents the loss calculated during the feature propagation process, represents the loss calculated during the label propagation process. The formula means to calculate the loss between the true label and the predicted label during the label propagation and feature propagation processes respectively, and obtain the custom loss function through the summation operation. Then, the learning optimizer is used to minimize the loss, and the optimizer is used to correct the graph convolutional neural network with an embedded hierarchical attention mechanism to improve the accuracy of scholar label prediction.

[0034] A user profile construction system based on a scientific research network, characterized by including:

[0035] (1) Data preprocessing module: Based on the co-authorship data information of papers, a scientific research social network graph is constructed with scholars as nodes and the co-authorship relationship between scholars as edges. The node features are processed into feature vectors, and the labels of the known nodes are converted into label vectors;

[0036] (2) Feature propagation module: The scientific research social network graph structure and node features are used as the input of the feature propagation layer. The feature propagation layer embeds a hierarchical attention mechanism in the graph convolutional neural network. A layer of multi-head attention mechanism is used to calculate the attention similarity coefficients for adjacent paired nodes respectively. After normalizing the attention similarity coefficients, a probability transition matrix is obtained. The probability transition matrix is linearly transformed with the original feature vector of the scholar to obtain the updated scholar feature. Finally, a layer of single-head attention is used to fuse the updated scholar features and perform label classification to predict the labels of the active research fields of the nodes;

[0037] (3) Label Propagation Module: The structure of the scientific research social network graph and the label vectors of some known scholars are used as the input of the label propagation layer. The label propagation layer refers to the label propagation algorithm, and the propagation of labels is achieved through the label propagation algorithm.

[0038] (4) Label Prediction Module of User Portrait: According to the sum of the losses calculated by the feature propagation layer and the label propagation layer, a customized loss function is obtained and minimized to help the feature propagation layer learn the optimizer. The optimizer is fed back to the feature propagation layer to learn the final node representation, and the final output result is the predicted research field label of the scholar.

[0039] The implementation process of the system is the same as that of the method implementation process.

[0040] A computer device, which includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the above computer program, the method for constructing a user portrait based on a scientific research network as described above is implemented.

[0041] A computer-readable storage medium stores a computer program for executing the method for constructing a user portrait based on a scientific research network as described above.

[0042] Compared with the prior art, the beneficial effects of the present invention are as follows: The method for constructing a user portrait for a scientific research network realizes the label prediction of unknown scholars by embedding an attention module in a graph convolutional neural network and integrating it with a label propagation algorithm. This method realizes the label prediction of the central node by fusing the feature and label information of adjacent nodes; the self-attention mechanism can aggregate the co-author information of papers according to the association structure between nodes and adaptively match the weights of different scholar (node) relationships; the label propagation algorithm can extract the association structure between nodes and increase the edge weights between associated nodes; embedding a self-attention module in the GCN and integrating it with the LPA algorithm can fully mine the feature, label information of scholars and the association structure information between scholars, complete the label information of missing and unknown scholars, obtain more accurate user labels, and thus improve the accuracy of the user portrait and perfect the construction of the user portrait. Description of the Drawings

[0043] Figure 1 It is a flowchart of the method for constructing a user portrait based on a scientific research network provided by an embodiment of the present invention;

[0044] Figure 2 It is a flowchart of the feature propagation method of an embodiment of the present invention;

[0045] Figure 3 It is a schematic structural diagram of a scientific research social network of an embodiment of the present invention;

[0046] Figure 4 The flowchart of the label propagation method in the embodiments of the present invention. Specific embodiments

[0047] The present invention will be further clarified below in conjunction with specific embodiments. It should be understood that these embodiments are only used to illustrate the present invention and not to limit the scope of the present invention. After reading the present invention, those skilled in the art will fall within the scope defined by the appended claims of this application for various equivalent forms of modification of the present invention.

[0048] As Figure 1 shown, the flowchart of the method for constructing a user portrait based on a scientific research network, and its working process is described as follows:

[0049] 1) Based on the data information of the co-authors of scientific research scholars, taking scholars as nodes and the co-authorship relationship between nodes as edges, construct a scientific research social network graph that can quantify the cooperation relationship between scholars. When constructing the scientific research social network graph, use one-hot encoding to process the node features into feature vectors, use csr_matrix to compress the eigenvalue of the scholars, and construct a sparse matrix; use one-hot encoding to process the existing node labels into label vectors, and use the csr_matrix or networks method to construct an adjacency matrix according to the association between nodes, that is, draw a scientific research social network graph based on the relationship between scholars. The scientific research social network can be expressed as G=(V, A, X, L), where G is the name of the undirected graph composed of nodes and the cooperation relationship between nodes, V is the node set composed of nodes, A represents the adjacency matrix constructed based on the association between users, A={a i1 ,a i2 ,a i3 ,···a ij}; X represents the node feature set, L represents the label set, and only some nodes have labels. The label set is L={1,···,c}. In this embodiment, the dataset for constructing the scientific research social network mainly refers to the co-author dataset (Co-auther Dataset), which is divided into a training set, a validation set, and a test set. The nodes represent scholars, the node features represent the paper keywords of each scholar, and the labels represent the most active research fields of each scholar;

[0050] 2) Feature Propagation: The structure of the scientific research social network graph and node features are used as the input of the feature propagation layer. The feature propagation layer embeds a hierarchical attention mechanism in the graph convolutional neural network. The specific operation is to calculate the attention similarity coefficients for adjacent paired nodes respectively using a layer of multi-head attention mechanism, normalize the attention similarity coefficients to obtain a probability transition matrix, linearly transform the probability transition matrix with the original feature vector of the scholar to obtain the updated scholar features, and finally fuse the updated scholar features using a layer of single-head attention and perform label classification using the softmax activation function to predict the label of the active research field of the node.

[0051] 3) Label Propagation: The structure of the scientific research social network graph and the label vectors of some known scholars are used as the input of the label propagation layer. The label propagation layer refers to using the label propagation algorithm to achieve label propagation. During the label propagation process, the labels of the labeled data are kept unchanged, and the labels are passed to the unlabeled data through the propagation matrix. Eventually, the label propagation of the user nodes tends to be stable. At this time, the loss of label propagation is calculated through the cross-entropy loss function.

[0052] 4) Label Prediction of User Portrait: Calculate the sum of the losses of the feature propagation layer and the label propagation layer to obtain a custom loss function, minimize the loss to process the learning optimizer, and feedback the custom loss function and the optimizer to the GCN with a hierarchical attention mechanism to learn the feature representation of the scholar. The final output result is the predicted label of the scholar's research field.

[0053] Figure 2 It is the flowchart of the feature propagation of this embodiment. A hierarchical attention mechanism is added to the graph convolutional neural network structure. First, use the multi-head attention mechanism layer to aggregate the features of neighbor nodes to achieve adaptive matching of weights according to different cooperation relationships, obtain attention coefficients, and obtain a probability transition matrix through normalization operations; then multiply the original feature vector of each node by the probability transition matrix to calculate the updated feature information of each node; finally, use a layer of single-head attention mechanism to fuse the updated node features obtained, and perform label classification using the softmax activation function to predict the label of the active research field of the node. The feature propagation further includes the following steps:

[0054] Step:1: Use the scientific research social network obtained after data preprocessing and a set of node feature vectors as the input of the feature propagation layer. The set of node feature vectors is Where N is the number of nodes and F is the dimension of node features. Set the hyperparameters of the graph convolutional neural network, including batch size (batch_size), number of epochs (epochs), patience (patience), learning rate (lr), weight loss (l2_coef), number of attention heads (hid_units), and number of attentions in the last layer (n_heads);

[0055] Step2: Train the multi-head attention mechanism: Use the multi-head attention mechanism to calculate the attention coefficients between the central node and its adjacent nodes respectively. As shown in the appendix, for node i, the adjacent nodes are scholar u Figure 3 as shown, for node i, the adjacent nodes are scholar u 1 、scholar u 2 、scholar u 3 and scholar u 4 . Quantify the importance of adjacent scholars to the central node through the attention mechanism. Among them, the importance of scholar i to scholar u 1 is different from the importance of scholar u 1 to scholar i. The attention coefficient α ij between node i and node j is calculated as follows:

[0056]

[0057] where W represents the trainable weight matrix, W ∈ R F ' ×F . Perform at least one linear transformation on the input feature vector through the weight matrix to obtain the output feature dimension. represents the parameter of the attention mechanism. The formula represents concatenating the linearly transformed input features of two adjacent nodes, and then performing a dot product operation with the attention parameter to obtain the attention coefficient;

[0058] Step3: Calculate the probability transition matrix: Normalize the attention coefficients through softmax, and use μ ij = softmax(α ij ) to obtain the probability transition matrix μ ij between node i and node j;

[0059] Step4: Feature propagation: Since there are many node features in the scientific research network social graph, in order to improve the calculation efficiency, use the multi-head attention mechanism for feature propagation. The essence of multi-head attention is to concurrently assign weights to neighbor nodes. Combine the feature vector of the node with the probability transition matrix calculated through the multi-head attention mechanism, that is, the updated feature of the central node is the weighted average of the features of adjacent scholar nodes. represents the updated feature of node i, and the calculation formula is:

[0060]

[0061] Among them, represents the probability transition matrix obtained by the k-th head attention calculation, where ij, j ∈ N(i) represent node i and its neighbor nodes, represents the feature vector of neighbor node j, and σ(·) uses the activation function Relu; the formula means that first, the partial updated features of node i are obtained through the multi-head attention mechanism, and then the partial updated features obtained by the k-head attention mechanism are concatenated to obtain the updated features of node i;

[0062] Step5: Through a single-head attention mechanism, fuse the updated features of each node and output a set of updated feature vectors, that is, output the updated features according to the input N node features. Suppose the node feature dimension of this new predicted feature vector is F’, and the output features can be expressed as

[0063] Step6: Label prediction: Use the softmax activation function to predict the research field labels corresponding to the node features, calculate the loss between the predicted feature values and the true values in the training set using the cross-entropy function, and use the loss function and the gradient descent function to optimize the feature propagation process to achieve the scholar research field label prediction task.

[0064] Figure 4 This is the flowchart of label propagation in the method of the present invention. The basic assumption of LPA is that connected nodes may have the same label, that is, connected scholars may have the same research field, and it further includes the following content:

[0065] Step1: Input: Take the labeled partial nodes, unlabeled nodes, and adjacency matrix A in the scientific research social network graph G as the input of the label propagation process. When initializing, set the labeled nodes as one-hot vectors and the unlabeled nodes as zero vectors;

[0066] Step2: Iterative label propagation algorithm: Let be the label matrix in the k-th iteration, where the element l i (k) represents the prediction distribution of the scholar labels in the k-th iteration. The calculation formula for label propagation after k iterations is as follows:

[0067] L k+1 = D -1 AL (k)

[0068] Among them, A represents the adjacency matrix, D is the diagonal matrix, and the value is obtained by summing the elements in the corresponding row of the adjacency matrix A. The formula represents obtaining the propagation probability matrix according to the association degree between users. Each user node adds the label values propagated by the surrounding user nodes according to the probability matrix by weight and updates them to its own probability distribution;

[0069] Step3: Retain the labeled nodes: Since LPA hopes to retain the label information of known scholars, the formula is as follows:

[0070]

[0071] Among them, represents the original label. The formula means that after the propagation is completed, the labels of the known user nodes need to be set to the original label again;

[0072] Step4: Label prediction: Obtain a stable label propagation result through multiple iterations of the above Step2 and 3;

[0073] Step5: Calculate the loss of the label propagation layer: The final label in LPA is the weighted average of the neighbor nodes. Use the cross-entropy loss function in tensorflow to calculate the loss (loss) of the label propagation, that is, the error between the true label and the predicted label.

[0074] Label prediction of the user profile: Use label propagation to extract the association structure information between nodes, and feedback the result to the feature propagation layer to improve the accuracy of label prediction. It further includes the following content:

[0075] Extract the features, labels and association structure information of adjacent nodes through feature propagation and label propagation, and perform label prediction of the research fields of nodes. The calculation formula of the custom loss function loss is as follows:

[0076]

[0077] Among them, represents calculating the loss during the feature propagation process, represents calculating the loss during the label propagation process. The formula means calculating the losses between the true label and the predicted label during the label propagation and feature propagation processes respectively, and obtaining the custom loss function through the summation operation. Then minimize the loss to learn the optimizer, and use the optimizer to correct the graph convolutional neural network with the embedded hierarchical attention mechanism to improve the accuracy of scholar label prediction.

[0078] According to the above embodiments, the present invention realizes a method for constructing a user portrait for a scientific research network, embeds an attention module in a graph convolutional neural network and fuses it with a label propagation algorithm to realize label prediction for unknown nodes. This method makes full use of the node features, label information, and the associated structure information between nodes, calculates the weights of different scholar relationships through a hierarchical attention mechanism, and complements the label information of missing and unknown nodes based on the features and labels of adjacent users, improving the accuracy of user label prediction, and thus realizing the precise construction of user portraits.

[0079] A user portrait construction system based on a scientific research network, characterized by comprising:

[0080] (1) Data preprocessing module: Based on the co-authorship data information of papers, taking scholars as nodes and the co-authorship relationship between scholars as edges, constructing a scientific research social network graph, processing the node features into feature vectors, and converting the labels of known nodes into label vectors;

[0081] (2) Feature propagation module: Taking the scientific research social network graph structure and node features as the input of the feature propagation layer, the feature propagation layer embeds a hierarchical attention mechanism in a graph convolutional neural network, calculates the attention similarity coefficients for adjacent paired nodes respectively using a layer of multi-head attention mechanism, normalizes the attention similarity coefficients to obtain a probability transition matrix, linearly transforms the probability transition matrix with the original scholar feature vectors to obtain updated scholar features, and finally fuses the updated scholar features using a layer of single-head attention and performs label classification to predict the labels of the active research fields of the nodes;

[0082] (3) Label propagation module: Taking the scientific research social network graph structure and the label vectors of some known scholars as the input of the label propagation layer, the label propagation layer refers to the label propagation algorithm, and realizes the propagation of labels through the label propagation algorithm;

[0083] (4) User portrait label prediction module: Calculate the sum of the losses of the feature propagation layer and the label propagation layer to obtain a custom loss function, minimize it to help the feature propagation layer learn the optimizer, feedback the optimizer to the feature propagation layer to learn the final node representation, and the final output result is the predicted research field label of the scholar.

[0084] Obviously, those skilled in the art should understand that each step of the method for constructing a user portrait based on a scientific research network in the above embodiments of the present invention or each module of the system for constructing a user portrait based on a scientific research network can be implemented by a general-purpose computing device. They can be concentrated on a single computing device or distributed on a network composed of multiple computing devices. Optionally, they can be implemented by program codes executable by the computing device. Thus, they can be stored in a storage device and executed by the computing device. And in some cases, the steps shown or described can be executed in a sequence different from that here, or they can be separately fabricated into individual integrated circuit modules, or multiple modules or steps among them can be fabricated into a single integrated circuit module for implementation. In this way, the embodiments of the present invention are not limited to any specific combination of hardware and software.

Claims

1. A method for constructing a user portrait based on a scientific research network, characterized in that, it includes the following steps: (1) Data preprocessing: Based on the co-authorship data information of papers, taking scholars as nodes and the co-authorship relationship between scholars as edges, construct a scientific research social network graph, process the node features into feature vectors, and convert the labels of known nodes into label vectors; (2) Feature propagation: Take the scientific research social network graph structure and node features as the input of the feature propagation layer. The feature propagation layer embeds a hierarchical attention mechanism in a graph convolutional neural network. Use a layer of multi-head attention mechanism to calculate the attention similarity coefficient for adjacent paired nodes respectively. After normalizing the attention similarity coefficient, obtain a probability transition matrix. Linearly transform the probability transition matrix with the original feature vector of the scholar to obtain the updated scholar feature. Finally, use a layer of single-head attention to fuse the updated scholar features and perform label classification to predict the label of the active research field of the node; (3) Label propagation: Take the scientific research social network graph structure and the label vectors of some known scholars as the input of the label propagation layer. The label propagation layer refers to the label propagation algorithm, and realizes the propagation of labels through the label propagation algorithm; (4) Label prediction of the user portrait: Calculate the sum of the losses of the feature propagation layer and the label propagation layer to obtain a custom loss function, minimize it to help the feature propagation layer learn the optimizer, and feedback the optimizer to the feature propagation layer to learn the final node representation. The final output result is the predicted research field label of the scholar; During the label propagation process, the basic assumption of the label propagation algorithm is that connected nodes may have the same label, that is, connected scholars may have the same research direction or research interest. Therefore, the label can be propagated iteratively along the edge to supplement the missing label information of the node. The specific content includes the following: 3.1) Model input: Take the label vector L and the adjacency matrix A in the scientific research social network graph G as the input of the label propagation process. When initializing, set the labeled nodes as one-hot vectors and the unlabeled nodes as zero vectors; 3.2) Label Propagation: Let be the label matrix in the k-th iteration, where the iteration refers to the process of continuously repeating label propagation, and the element l i (k) represents the predicted distribution of node labels in the k-th iteration. The calculation formula for label propagation after the k-th iteration is as follows: L k+1 = D -1 AL (k) Among them, A represents the adjacency matrix, D is the diagonal matrix, and the value is obtained by summing the corresponding row elements in the adjacency matrix A. The formula represents the label transfer along the edge; 3.3) Retain the labeled nodes: Because the label information of some known scholars is determined, after the propagation is completed, the labels of the known user nodes need to be set back to the original labels; 3.4) The final label in LPA is the weighted average of the neighbor nodes, and the cross-entropy loss function in tensorflow is used to calculate the loss of label propagation.

2. The method for constructing a user portrait based on a scientific research network according to claim 1, characterized in that, During the label propagation process, keep the labels of the labeled data unchanged, and let it pass the labels to the unlabeled data through the propagation matrix; finally, the label propagation of the node tends to be stable, and at this time, the cross-entropy loss function is used to calculate the loss of label propagation.

3. The method for constructing a user portrait based on a scientific research network according to claim 1, characterized in that, In the above data preprocessing, based on the co-authorship data information of papers, a scientific research social network graph that can quantify the cooperation relationship of scientific research users is constructed with scholars as nodes and the co-authorship relationship between scholars as edges. The scientific research social network graph is represented as G=(V, A, X, L), where G represents the name of the scientific research social network graph, V represents the node set, representing scholars; A represents the adjacency matrix, representing the co-authorship relationship between scholars; X represents the feature set of scholars, that is, the node feature set, and the node feature represents the keywords of each author's paper. The construction method of the feature vector is: using one-hot encoding to mark the included features as 1 and the non-included features as 0; L represents the label set owned by some scholars, and the label represents the label of the most active research field of the scholar. The construction method of the label vector is: using one-hot encoding to mark the appeared labels as 1 and the non-appeared labels as 0.

4. The method for constructing a user portrait based on a scientific research network according to claim 1, characterized in that, during the feature propagation process, a hierarchical attention mechanism is added to the structure of the graph convolutional neural network (GCN). First, the multi-head attention mechanism layer is used to aggregate the features of neighbor nodes to achieve adaptive matching of weights according to different cooperation relationships, obtain the attention coefficients, and obtain the probability transition matrix through normalization operations; then, the original feature vectors of each node are multiplied by the probability transition matrix to calculate the updated feature information of each node; finally, a single-head attention mechanism is used to fuse the updated node features obtained, and the softmax activation function is used for label classification to predict the label of the active research field of the node.

5. The method for constructing a user portrait based on a scientific research network according to claim 1, characterized in that, the feature propagation further includes the following steps: 2.1) Input: The scientific research social network graph and a set of node feature vectors are used as the input of the feature propagation layer, and the set of feature vectors of scholars is where N is the number of nodes and F is the dimension of node features; 2.2) Calculate the probability transition matrix: Taking a certain node as the center, calculate the attention similarity coefficient of its adjacent nodes. Then this node is called the central node, and the adjacent nodes are called neighbor nodes. Use the multi-head attention mechanism to calculate the attention similarity coefficient between each central node and its first-order adjacent nodes, and perform a normalization operation through the softmax activation function to obtain the probability transition matrix μ ij ; 2.3) Feature propagation: Use the multi-head attention mechanism for feature propagation. The essence of the multi-head attention is to concurrently execute the calculation of the probability weights between neighbors, and combine the feature vector of the node with the probability transition matrix calculated by the multi-head attention mechanism, that is, the updated feature of the central node is the weighted average of the features of adjacent nodes; first, the partial updated feature of node i is obtained through the multi-head attention mechanism, and then the partial updated features obtained by the k-head attention mechanism are concatenated to obtain the updated feature of node i; 2.4) Fuse the updated features of each node through a single-headed attention network layer, and output a set of updated feature vectors. Update features are output according to the input N node features. Let the node feature dimension of this new predicted feature vector be F’, which is expressed as 2.5) Label prediction: Use the softmax activation function to predict the research field label corresponding to the node feature, calculate the error loss between the predicted label and the true label using the cross-entropy function, and use the loss function and the gradient descent function to optimize the feature propagation process to achieve the scholar research field label prediction task.

6. The method for constructing a user portrait based on a scientific research network according to claim 1, characterized in that, in the label prediction of the user portrait, the loss functions of feature propagation and label propagation are summed to construct a custom loss function, and the loss function is used to optimize the label prediction result of feature propagation to improve the accuracy of label prediction, including the following contents; Extract the features, labels and associated structure information of adjacent nodes through feature propagation and label propagation, and perform the research field label prediction of scholars. The calculation formula of the custom loss function loss is as follows: Among them, represents the loss during the calculation of feature propagation, represents the loss during the calculation of label propagation. The formula means to calculate the losses between the true labels and the predicted labels during label propagation and feature propagation respectively, and obtain a custom loss function through a summation operation; then minimize the loss to learn an optimizer, and use the optimizer to correct the graph convolutional neural network with an embedded hierarchical attention mechanism to improve the accuracy of scholar label prediction.

7. A user portrait construction system based on a scientific research network, characterized in that, comprising: (1) Data preprocessing module: Based on the co-authorship data information of papers, taking scholars as nodes and the co-authorship relationship between scholars as edges, constructing a scientific research social network graph, processing the node features into feature vectors, and converting the labels of known nodes into label vectors; (2) Feature propagation module: Taking the scientific research social network graph structure and node features as the input of the feature propagation layer. The feature propagation layer embeds a hierarchical attention mechanism in a graph convolutional neural network, uses a layer of multi-head attention mechanism to calculate the attention similarity coefficient for adjacent paired nodes respectively, normalizes the attention similarity coefficient to obtain a probability transition matrix, linearly transforms the probability transition matrix with the original feature vector of the scholar to obtain the updated scholar feature, and finally uses a layer of single-head attention to fuse the updated scholar feature and perform label classification to predict the label of the active research field of the node; (3) Label propagation module: Taking the scientific research social network graph structure and the label vectors of some known scholars as the input of the label propagation layer. The label propagation layer refers to the label propagation algorithm to achieve the propagation of labels; (4) User portrait label prediction module: Calculate the sum of the losses of the feature propagation layer and the label propagation layer to obtain a custom loss function, minimize it to help the feature propagation layer learn the optimizer, and feedback the optimizer to the feature propagation layer to learn the final node representation. The final output result is the predicted research field label of the scholar; During the label propagation process, the basic assumption of the label propagation algorithm is that connected nodes may have the same label, that is, connected scholars may have the same research direction or research interest. Therefore, the label can be propagated iteratively along the edge to supplement the missing label information of the node. The specific content includes the following: 3.1) Model input: Taking the label vector L and the adjacency matrix A in the scientific research social network graph G as the input of the label propagation process. At initialization, the labeled nodes are set as one-hot vectors, and the unlabeled nodes are set as zero vectors; 3.2) Label Propagation: Let be the label matrix in the k-th iteration, where the iteration refers to the process of continuously repeating label propagation, and the element l i (k) represents the predicted distribution of node labels in the k-th iteration. The calculation formula for label propagation after the k-th iteration is as follows: L k+1 = D -1 AL (k) where, A represents the adjacency matrix, D is a diagonal matrix, and the value is obtained by summing the corresponding row elements in the adjacency matrix A. The formula represents the label transfer along the edge; 3.3) Retain the labeled nodes: Since the label information of some known scholars is determined, after the propagation is completed, the labels of the known user nodes need to be reset to the original labels; 3.4) The final label in LPA is the weighted average of the neighbor nodes, and the cross-entropy loss function in tensorflow is used to calculate the loss of label propagation.

8. A computer device, characterized in that: The computer device includes a memory, a processor, and a computer program stored on the memory and executable on the processor. When the processor executes the above computer program, it implements the method for constructing a user portrait based on a scientific research network according to any one of claims 1-6.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program for executing the method for constructing a user profile based on a scientific research network according to any one of claims 1-6.

Images