Academic network node classification method and apparatus

Abstract

The application discloses an academic network node classification method and device, and the method comprises the following steps: constructing an undirected graph network based on an academic network; establishing an optimization objective function by using a preset feature transformation matrix and a low-rank matrix, an adjacency matrix and a feature matrix; solving gradient information of the feature transformation matrix and the low-rank matrix by using the optimization objective function, updating matrix parameters of the feature transformation matrix and the low-rank matrix based on the gradient information to obtain updated matrix parameters; obtaining a classification prediction probability matrix of the feature transformation matrix and the low-rank matrix for academic entity nodes according to the updated matrix parameters, and obtaining a classification prediction result of the academic network node based on the classification prediction probability matrix. The application can realize reasonable classification of academic network nodes in a noisy condition.

Description

Technical Field

[0001] The present invention relates to the technical field of node classification, and in particular to a method and device for classifying nodes in an academic network. Background Art

[0002] An academic network is a network structure that reflects the academic exchanges and collaborations among scholars. Nodes in this network typically represent various academic entities, such as researchers, papers, academic journals, and academic institutions. Edges represent the various academic relationships between these entities, such as collaborations, citations, and publications. Academic networks serve as a crucial data foundation for research in areas such as academic information retrieval, academic recommendation, and academic evaluation. Node classification on academic networks primarily involves categorizing these academic entities. For example, researchers can be classified based on their research areas, or papers can be categorized by topic based on their content. Node classification in academic networks is a very challenging problem, as the categories of academic entities are diverse and may overlap. Furthermore, the structural characteristics of academic networks, such as community structures and core-periphery structures, also pose challenges to node classification.

[0003] Spectral Graph Neural Networks (SGNNs) are a mainstream tool for node classification in academic networks. SGNNs consist of multiple graph convolutional transformation layers, each of which includes a feature transformation step and a spectral graph convolution step. The former, typically implemented by fully connected network layers, aims to extract feature representations for each node in the graph; the latter, based on the graph structure, extracts cross-node information for downstream tasks. Spectral graph convolutions first define a graph spectral domain transformation (i.e., graph Fourier transform) by leveraging the eigenvector matrix of the graph (i.e., academic network) adjacency matrix. Then, an N-dimensional diagonal matrix is ​​introduced in the spectral domain to form a graph spectral filter, where N represents the number of nodes in the academic network. However, computing the eigenvectors of the adjacency matrix incurs significant computational overhead. To overcome this computational bottleneck, many subsequent works have proposed directly approximating the graph spectral filter by utilizing various polynomial bases (e.g., Chebyshev polynomials and Bernstein polynomials).

[0004] Based on the above ideas, GCN [Kipf, 2017] uses the Chebyshev polynomials of the adjacency matrix as a basis to construct graph filters and improves computational efficiency by truncating high-order terms. Wu et al. [Wu, 2019] proposed simplifying GCN by eliminating nonlinear activations in intermediate network layers, further reducing computational overhead without sacrificing basic performance. Li et al. [Li, 2018] showed that convolution in GCN is actually a special form of Laplace smoothing. If the given graph contains sufficient information, it can filter out noise signals from the original features. Conversely, when a given graph contains too many noisy links, multiple graph convolutions will cause incorrect feature aggregation, leading to the problem of oversmoothing. GCNII [Chen, 2020] extends GCN by introducing initial residuals and identity mapping, which has been shown to effectively alleviate the oversmoothing problem. GraphHeat [Xu, 2019] uses a heat kernel formula to assign greater importance to low-frequency filters, thereby reducing the impact of high-frequency interference on graph convolution. BernNet [He, 2021] proposes replacing Chebyshev polynomials with Bernstein polynomials, which have been shown to provide better approximations of ideal filters. Under ideal conditions, this polynomial-based construction can approximate theoretically optimal graph filters. However, in real-world scenarios, the adjacency matrix of academic networks can contain significant noise, disrupting these ideal conditions and resulting in poor node classification performance. Summary of the Invention

[0005] The present invention aims to solve one of the technical problems in the related art at least to a certain extent.

[0006] To this end, this paper proposes a method for classifying academic network nodes. This method goes beyond the spectral transformation construction based on adjacency matrix polynomials and directly learns a low-rank representation of the spectral projection matrix. This method also ensures learning effectiveness by incorporating orthogonality and graph smoothing regularization. This method produces a more optimized spectral transformation matrix than traditional methods, significantly improving the accuracy of academic network node classification.

[0007] Another object of the present invention is to provide a device for classifying academic network nodes.

[0008] To achieve the above objectives, the present invention proposes a method for classifying academic network nodes, comprising:

[0009] Constructing an undirected graph network based on the academic network; wherein the undirected graph network includes a set of academic entity nodes, relationships between different academic entities, an adjacency matrix of the academic network, and a feature matrix of the academic network nodes;

[0010] Establishing an optimization objective function using a preset feature transformation matrix and a low-rank matrix as well as the adjacency matrix and the feature matrix;

[0011] Using the optimization objective function to respectively solve the gradient information of the feature transformation matrix and the low-rank matrix, and updating the matrix parameters of the feature transformation matrix and the low-rank matrix based on the gradient information to obtain updated matrix parameters;

[0012] According to the updated matrix parameters, the classification prediction probability matrix of the feature transformation matrix and the low-rank matrix for the academic entity node is obtained, and the classification prediction result of the academic network node is obtained based on the classification prediction probability matrix.

[0013] The method for classifying academic network nodes according to the embodiment of the present invention may also have the following additional technical features:

[0014] In one embodiment of the present invention, before establishing the optimization objective function, the method further includes:

[0015] Get the degree of academic entity node;

[0016] Calculating the degree of the academic entity node to obtain a node degree calculation result;

[0017] The adjacency matrix and the feature matrix are preprocessed based on the node degree calculation result to obtain a preprocessed adjacency matrix and feature matrix.

[0018] In one embodiment of the present invention, the degree of the academic entity node is the number of adjacent nodes of each academic entity node; the academic entity includes papers and / or authors.

[0019] In one embodiment of the present invention, the preprocessing of the adjacency matrix and the feature matrix based on the node degree calculation result to obtain the preprocessed adjacency matrix and feature matrix includes:

[0020] Calculate the degree of each academic entity node and construct a diagonal matrix using the degree of the academic entity node as a diagonal element;

[0021] Performing a square root operation on the diagonal matrix to obtain an inverse operation diagonal matrix, and performing a normalization operation on the inverse operation diagonal matrix and the adjacency matrix to obtain a normalized adjacency matrix; and

[0022] Each column of the feature matrix is ​​normalized using the ZScore method to obtain a normalized feature matrix.

[0023] In one embodiment of the present invention, the matrix parameters of the feature transformation matrix are updated using gradient descent; the matrix parameters of the low-rank matrix are updated using the Riemann gradient projection method for manifold optimization; and the classification prediction results of the academic network nodes based on the classification prediction probability matrix are obtained using a greedy selection algorithm.

[0024] To achieve the above-mentioned purpose, the present invention further provides a device for classifying nodes in an academic network, comprising:

[0025] An undirected graph network construction module is used to construct an undirected graph network based on an academic network; wherein the undirected graph network includes a set of academic entity nodes, the relationships between different academic entities, the adjacency matrix of the academic network, and the feature matrix of the academic network nodes;

[0026] An objective function construction module is used to establish an optimization objective function using a preset feature transformation matrix and a low-rank matrix as well as the adjacency matrix and the feature matrix;

[0027] A matrix parameter updating module is used to use the optimization objective function to solve the gradient information of the feature transformation matrix and the low-rank matrix respectively, and update the matrix parameters of the feature transformation matrix and the low-rank matrix based on the gradient information to obtain updated matrix parameters;

[0028] The node classification prediction module is used to obtain the classification prediction probability matrix of the feature transformation matrix and the low-rank matrix for the academic entity node according to the updated matrix parameters, and obtain the classification prediction result of the academic network node based on the classification prediction probability matrix.

[0029] The academic network node classification method and device of the embodiment of the present invention breaks away from the spectral transformation construction method based on adjacency matrix polynomials, directly learns the low-rank expression of the spectral projection matrix, and ensures the learning effect by adding orthogonality constraints and graph smoothing regularization, thereby realizing reasonable classification of academic network nodes in the presence of noise.

[0030] Additional aspects and advantages of the present invention will be set forth in part in the description which follows and, in part, will be obvious from the description which follows, or may be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments in conjunction with the accompanying drawings, in which:

[0032] Figure 1 is a flow chart of a method for classifying academic network nodes according to an embodiment of the present invention;

[0033] Figure 2is a flow chart of another method for classifying academic network nodes according to an embodiment of the present invention;

[0034] Figure 3 This is a typical schematic diagram of academic network node classification according to an embodiment of the present invention;

[0035] Figure 4 2 is a schematic diagram of the structure of an academic network node classification device according to an embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments and features of the embodiments of the present invention can be combined with each other. The present invention will be described in detail below with reference to the accompanying drawings and in combination with the embodiments.

[0037] In order to enable those skilled in the art to better understand the solutions of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the embodiments described 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 making creative efforts should fall within the scope of protection of the present invention.

[0038] The following describes the academic network node classification method and device proposed according to an embodiment of the present invention with reference to the accompanying drawings.

[0039] Figure 1 This is a flowchart of a method for classifying academic network nodes according to an embodiment of the present invention.

[0040] like Figure 1 As shown, the method includes but is not limited to the following steps:

[0041] S1, construct an undirected graph network based on the academic network; the undirected graph network includes a set of academic entity nodes, the relationships between different academic entities, the adjacency matrix of the academic network, and the feature matrix of the academic network nodes;

[0042] S2, using the preset feature transformation matrix and low-rank matrix as well as the adjacency matrix and feature matrix to establish the optimization objective function;

[0043] S3, using the optimization objective function to solve the gradient information of the feature transformation matrix and the low-rank matrix respectively, and updating the matrix parameters of the feature transformation matrix and the low-rank matrix based on the gradient information to obtain updated matrix parameters;

[0044] S4, obtains the classification prediction probability matrix of the academic entity node based on the feature transformation matrix and the low-rank matrix according to the updated matrix parameters, and obtains the classification prediction result of the academic network node based on the classification prediction probability matrix.

[0045] In one embodiment of the present invention, a given academic network can be represented by an undirected graph network G = (V, E, A, X), where V represents a set of |V| academic entity nodes (e.g., authors, papers), and E represents a set of |E| edges, representing the relationships between different academic entities (e.g., whether there is collaboration between different authors, or whether there is a citation relationship between different papers). is the matrix of E (i.e. the adjacency matrix of the academic network), Represents the feature matrix of |V| academic network nodes. The output is a one-hot matrix Indicates which of the c categories |V| nodes belong to (for example, the author or paper belongs to a discipline such as medicine, physics, or computer science).

[0046] For the above problem, the traditional spectral neural network approach is to first normalize the angle matrix D of A to obtain It has better spectral transformation properties. Then use Polynomial Construct a graph filter to filter the academic network node feature X, and then apply a linear transformation to the filtered feature Finally, the final output is obtained through a nonlinear activation function σ:

[0047]

[0048] in The model outputs the academic network node classification matrix, where each row represents the probability distribution vector of the c categories to which the corresponding node belongs. For the sake of clarity, only the single-layer network case is considered here, and the multi-layer case is similar. The effectiveness of The quality of When no noise information is included, It will be able to effectively approximate the ideal filter However, when When it contains more noise information, will with There is always a non-trivial error ε between them, which leads to errors in the classification of academic network nodes.

[0049] In order to improve the accuracy of academic network node classification and break through the error bound of the above polynomial graph filter, this paper proposes to use a low-rank matrix replace Realize The approximation of is a "thin and tall" matrix, that is, |V|>>c. Note that the width of U is equal to the number of node classifications c. It can be theoretically proved that this setting can ensure UU T Perfect Approximation In addition, to improve the approximation effect in practice, orthogonal constraints and graph smoothing regularization are imposed on U. The feature linear transformation W and nonlinear activation function σ are further introduced to obtain the following optimization model:

[0050]

[0051] sU T U=I C

[0052] The first term in this formula Encourage the model to estimate the node classification σ(UU T XW) is close to the true labeled classification Y, the second Encourage UU T Capturing the main spectral structure of the academic network adjacency matrix, the constraint condition U T U=I C Avoid U from falling into a trivial solution. Note that there are two variables that need to be optimized and there are constraints in the above optimization problem. Gradient descent and Riemann gradient projection method will be used to optimize and update them respectively until the update iteration converges. After the optimization is completed, the model output σ(UU T XW) adopts the greedy selection principle to obtain the classification prediction of each academic network node, such as Figure 2 shown.

[0053] The following describes in detail the academic network node classification method according to an embodiment of the present invention with reference to the accompanying drawings.

[0054] Figure 3 is a typical academic network node classification diagram, such as Figure 3 As shown in the figure, there are 7 academic author nodes, and their academic connections can be represented by a 7×7 adjacency matrix A=[a ij ] characterization, where a ij =1 indicates node v i With v j There is an edge between them (there is academic cooperation), otherwise a ij = 0. In addition, each node has a corresponding author information feature ( Figure 3 (represented by a 4-dimensional vector in [ ] ), can be represented by a 7×4 matrix X. The category information for some author nodes is already given: v2 and v3 are category 1 (e.g., physics), and v5 and v6 are category 2 (e.g., computer science). The goal of the graph neural network method with learnable spectral projection is to infer the category information for the remaining author nodes.

[0055] In one embodiment of the present invention, the adjacency matrix and node feature preprocessing are performed. First, the degree of each author node is calculated, that is, the number of adjacent nodes of each node. Figure 3 For example, the degrees of nodes v1 to v7 are 2, 3, 3, 3, 2, 2, 3. These 7 values ​​are used as diagonal elements to construct a 7×7 diagonal matrix D. Taking the square root of D and taking the inverse operation, we can get D -0.5 , based on this, A is normalized This ensures The absolute value of the eigenvalue of will not exceed 1, which is beneficial to the stability of the algorithm. In addition, each column of the node feature matrix X is standardized using the ZScore method (that is, subtracting the mean of each column and dividing it by the standard deviation of the corresponding column), which is beneficial to the convergence of the algorithm.

[0056] In one embodiment of the present invention, the model and the objective function are defined based on the processed normalized adjacency matrix With the feature matrix X, the feature transformation matrix W is further introduced to perform feature transformation learning. Figure 3 For the example shown, the dimension of the node feature is 4 and the number of categories is 2, so the size of W is 4×2. In addition, the matrix U is introduced to approximate the ideal propagation matrix. Figure 3 In the example shown, the number of nodes is 7 and the number of categories is 2, so the size of U is 7 × 2. Based on these matrices, the following optimization objective function is established:

[0057]

[0058] sU T U=I2

[0059] The nonlinear activation σ is the softmax function, Tr(·) represents the matrix trace (summing the diagonal elements), I2 represents the 2×2 identity matrix, and α is the regularization weight, typically set to 0.25. U and W are randomly initialized and then updated using an alternating optimization approach.

[0060] In one embodiment of the present invention, the feature transformation matrix W is updated. First, U is fixed and W is updated. Note that for W, this is an unconstrained optimization problem, so the update can be performed using gradient descent. First, the gradient information of the objective function with respect to W is calculated:

[0061]

[0062] Update W using the gradient descent method:

[0063]

[0064] Here γ1 is the update step size and is usually set to 0.2.

[0065] In one embodiment of the present invention, the low-rank matrix U is updated. In this stage, W is fixed and U is updated. Note that U has constraints, so it cannot be directly updated using gradient descent. In fact, the constraints limit U to a Grassmann manifold, so the Riemann gradient projection method for manifold optimization can be used to update U. Specifically, first calculate the normal gradient of the objective function with respect to U:

[0066]

[0067] where Z = σ′(UU T XW)⊙(σ(UU T XW)-Y). Based on Its Riemann gradient on the Grassmann manifold can be calculated:

[0068]

[0069] Then update U based on the Riemann gradient:

[0070] U←U-γ2grad[U]

[0071] Here the update step size γ2 is usually set to 0.2. The above update may destroy the Grassmann manifold condition, so it is necessary to project the updated U onto the Grassmann manifold to make it satisfy the constraint conditions:

[0072]

[0073] Among them S L With S R The left singular value decomposition vector matrix and the right singular value decomposition vector matrix of U respectively.

[0074] In one embodiment of the present invention, greedy selection is used to obtain node classification. Based on the optimized model parameters W and U, the classification prediction probability matrix of the model for the author node is obtained:

[0075]

[0076] for Figure 3 For example, The size of is 7 rows (number of nodes) and 2 columns (number of categories). In order to obtain the categories of v1, v4, and v7, The 1st, 4th, and 7th lines respectively select the category with the highest probability as the prediction result of the model.

[0077] The experimental results of the present invention show that: on four common academic network node classification public datasets, Cora, Citeseer, Pubmed, and Coauthor-Physics, the proposed method was compared with ten advanced graph neural network methods, namely GraphSAGE, GAT, SGC, APPNP, DGC, ChebNet, GCN, GraphHeat, FAGCN, and S2GC. The results show that the proposed method achieved the highest classification accuracy compared with all ten methods on the four datasets, and achieved a 2% improvement in accuracy compared with the second place on the Citeseer dataset (the gap between the second and third place was only 0.3%). The proposed method also had the smallest standard deviation across all datasets, which demonstrates the stability of the method.

[0078] Furthermore, a noise robustness experiment was conducted on a simulated graph academic network dataset: noise was injected into a standard dataset in steps of 10% from 0% to 70%, and the performance of different methods on the noisy dataset was tested. The results showed that the proposed method was more stable than ten state-of-the-art graph neural networks, achieving the highest improvement of around 10%, demonstrating its robustness to noise.

[0079] Furthermore, a visual comparison of the graph propagation matrix learned by the traditional spectral graph neural network method and the proposed method was performed on four datasets: Cora, Citeseer, Pubmed, and Coauthor-Physics. The results show that the graph propagation matrix learned by the proposed method can break through the shackles of traditional spectral graph neural networks and achieve the most ideal approximation of the ideal propagation matrix.

[0080] According to the academic network node classification method of the embodiment of the present invention, it breaks away from the spectral transformation construction method based on the adjacency matrix polynomial, directly learns the low-rank expression of the spectral projection matrix, and ensures the learning effect by adding orthogonality constraints and graph smoothing regularization, so as to achieve reasonable classification of academic network nodes in the presence of noise.

[0081] In order to implement the above embodiment, Figure 4 As shown, this embodiment also provides an academic network node classification device 10, which includes an undirected graph network construction module 100, an objective function construction module 200, a matrix parameter updating module 300 and a node classification prediction module 400.

[0082] An undirected graph network construction module 100 is used to construct an undirected graph network based on an academic network; wherein the undirected graph network includes a set of academic entity nodes, relationships between different academic entities, an adjacency matrix of the academic network, and a feature matrix of the academic network nodes;

[0083] An objective function construction module 200 is used to establish an optimization objective function using a preset feature transformation matrix and a low-rank matrix as well as an adjacency matrix and a feature matrix;

[0084] A matrix parameter updating module 300 is used to respectively solve the gradient information of the feature transformation matrix and the low-rank matrix using the optimization objective function, and update the matrix parameters of the feature transformation matrix and the low-rank matrix based on the gradient information to obtain updated matrix parameters;

[0085] The node classification prediction module 400 is used to obtain the classification prediction probability matrix of the feature transformation matrix and the low-rank matrix for the academic entity node according to the updated matrix parameters, and obtain the classification prediction result of the academic network node based on the classification prediction probability matrix.

[0086] Furthermore, before the objective function construction module 200, the apparatus further includes a data preprocessing module for:

[0087] Get the degree of academic entity node;

[0088] Calculate the degree of the academic entity node to obtain the node degree calculation result;

[0089] The adjacency matrix and the characteristic matrix are preprocessed based on the node degree calculation results to obtain the preprocessed adjacency matrix and the characteristic matrix.

[0090] Furthermore, the degree of an academic entity node is the number of adjacent nodes of each academic entity node; the academic entity includes papers and / or authors.

[0091] Furthermore, the data preprocessing module is also used to:

[0092] Calculate the degree of each academic entity node and construct a diagonal matrix using the degree of the academic entity node as a diagonal element;

[0093] Performing a square root operation on the diagonal matrix to obtain an inverse operation diagonal matrix, and performing a normalization operation on the inverse operation diagonal matrix and the adjacency matrix to obtain a normalized adjacency matrix; and

[0094] Each column of the feature matrix is ​​normalized using the ZScore method to obtain a normalized feature matrix.

[0095] Furthermore, the matrix parameters of the feature transformation matrix are updated using gradient descent; the matrix parameters of the low-rank matrix are updated using the Riemann gradient projection method for manifold optimization; and the classification prediction results of the academic network nodes based on the classification prediction probability matrix are obtained using a greedy selection algorithm.

[0096] According to the academic network node classification device of the embodiment of the present invention, it breaks away from the spectral transformation construction method based on the adjacency matrix polynomial, directly learns the low-rank expression of the spectral projection matrix, and ensures the learning effect by adding orthogonality constraints and graph smoothing regularization, thereby realizing reasonable classification of academic network nodes in the presence of noise.

[0097] In the description of this specification, the reference terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" mean that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representations of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described can be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art can combine and combine different embodiments or examples described in this specification and features of different embodiments or examples without contradiction.

[0098] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of the technical features being referred to. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one such feature. In the description of the present invention, "plurality" means at least two, such as two, three, etc., unless otherwise specifically defined.

Claims

1. A method for classifying nodes in academic networks, characterized in that: The following steps are involved: Constructing an undirected graph network based on the academic network; wherein the undirected graph network includes a set of academic entity nodes, relationships between different academic entities, an adjacency matrix of the academic network, and a feature matrix of the academic network nodes; Establishing an optimization objective function using a preset feature transformation matrix and a low-rank matrix as well as the adjacency matrix and the feature matrix; Using the optimization objective function to respectively solve the gradient information of the feature transformation matrix and the low-rank matrix, and updating the matrix parameters of the feature transformation matrix and the low-rank matrix based on the gradient information to obtain updated matrix parameters; Obtaining a classification prediction probability matrix of the feature transformation matrix and the low-rank matrix for the academic entity node according to the updated matrix parameters, and obtaining a classification prediction result of the academic network node based on the classification prediction probability matrix; Before establishing the optimization objective function, the method further includes: Get the degree of academic entity node; Calculating the degree of the academic entity node to obtain a node degree calculation result; Preprocessing the adjacency matrix and the characteristic matrix based on the node degree calculation result to obtain a preprocessed adjacency matrix and characteristic matrix; The degree of the academic entity node is the number of adjacent nodes of each academic entity node; the academic entities include papers, authors, academic journals and academic institutions; the relationships include cooperation relationships, citation relationships and publication relationships; The preprocessing of the adjacency matrix and the feature matrix based on the node degree calculation result to obtain the preprocessed adjacency matrix and feature matrix includes: Calculate the degree of each academic entity node and construct a diagonal matrix using the degree of the academic entity node as a diagonal element; Performing a square root operation on the diagonal matrix to obtain an inverse operation diagonal matrix, and performing a normalization operation on the inverse operation diagonal matrix and the adjacency matrix to obtain a normalized adjacency matrix; and Each column of the feature matrix is ​​normalized using the ZScore method to obtain a normalized feature matrix.

2. The method according to claim 1, characterized in that The matrix parameters of the feature transformation matrix are updated using gradient descent; the matrix parameters of the low-rank matrix are updated using the Riemann gradient projection method for manifold optimization; and the classification prediction results of the academic network nodes based on the classification prediction probability matrix are obtained using a greedy selection algorithm.

3. A device for classifying nodes in an academic network, characterized in that: include: An undirected graph network construction module is used to construct an undirected graph network based on an academic network; wherein the undirected graph network includes a set of academic entity nodes, the relationships between different academic entities, the adjacency matrix of the academic network, and the feature matrix of the academic network nodes; An objective function construction module is used to establish an optimization objective function using a preset feature transformation matrix and a low-rank matrix as well as the adjacency matrix and the feature matrix; A matrix parameter updating module is used to use the optimization objective function to solve the gradient information of the feature transformation matrix and the low-rank matrix respectively, and update the matrix parameters of the feature transformation matrix and the low-rank matrix based on the gradient information to obtain updated matrix parameters; A node classification prediction module is used to obtain a classification prediction probability matrix of the feature transformation matrix and the low-rank matrix for the academic entity node according to the updated matrix parameters, and obtain a classification prediction result of the academic network node based on the classification prediction probability matrix; Before the objective function construction module, the device further includes a data preprocessing module for: Get the degree of academic entity node; Calculating the degree of the academic entity node to obtain a node degree calculation result; Preprocessing the adjacency matrix and the characteristic matrix based on the node degree calculation result to obtain a preprocessed adjacency matrix and characteristic matrix; The degree of the academic entity node is the number of adjacent nodes of each academic entity node; the academic entities include papers, authors, academic journals and academic institutions; the relationships include cooperation relationships, citation relationships and publication relationships; The data preprocessing module is further used to: Calculate the degree of each academic entity node and construct a diagonal matrix using the degree of the academic entity node as a diagonal element; Performing a square root operation on the diagonal matrix to obtain an inverse operation diagonal matrix, and performing a normalization operation on the inverse operation diagonal matrix and the adjacency matrix to obtain a normalized adjacency matrix; and Each column of the feature matrix is ​​normalized using the ZScore method to obtain a normalized feature matrix.

4. The device according to claim 3, characterized in that The matrix parameters of the feature transformation matrix are updated using gradient descent; the matrix parameters of the low-rank matrix are updated using the Riemann gradient projection method for manifold optimization; and the classification prediction results of the academic network nodes based on the classification prediction probability matrix are obtained using a greedy selection algorithm.

Images