Academic paper reference anomaly detection method based on mutual information representation learning
Through mutual information representation learning from global and local perspectives and graph convolutional neural networks, we generate discriminative embedding representations, solve the problem of over-smoothing of node features in academic paper citation networks, achieve efficient and unsupervised abnormal citation detection, and improve the credibility and robustness of academic networks.
Patent Information
- Application Number
- CN202510917974.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-03
- Publication Date
- 2025-10-03
AI Technical Summary
Existing graph neural network methods are prone to over-smoothing of node features in academic paper citation networks, affecting the identification of abnormal links. In addition, their reliance on label information leads to high costs and serious false detection and missed detection phenomena.
A mutual information representation learning method is adopted. Through collaborative modeling from global and local perspectives, graph convolutional neural networks are used for node embedding representation. Combined with K-means clustering and mutual information maximization, a discriminative embedding representation is generated to identify abnormal citation behavior.
Under unsupervised learning conditions, it can efficiently and accurately identify abnormal citation behaviors in academic paper networks, improve the distinguishability and detection accuracy of node representation, reduce training costs, and is suitable for sparse network structures, assisting in the detection of academic misconduct.
Smart Images
Figure CN120744762A_ABST
Abstract
Description
Technical Field
[0001] The present invention belongs to the field of information technology, and in particular relates to a method for detecting anomaly in academic paper citations based on mutual information representation learning. Background Art
[0002] With the widespread adoption and development of graph-structured data in a variety of practical applications, including social networks, financial systems, bioinformatics, and scientific research management, graph-based data analysis techniques have gradually become a key research direction in artificial intelligence and data mining. Graph anomaly detection, a key task for identifying potentially anomalous nodes, edges, or subgraph structures in a graph, possesses significant theoretical and practical significance. Taking the citation network of scientific research papers as an example, each paper can be modeled as a node in the graph, and the citation relationships between different papers as directed edges between nodes, thereby constructing a semantically directed academic citation graph. Normally, citation edges typically reflect the inheritance of research topics, the borrowing of methods, or the evolution of academic perspectives, exhibiting reasonable academic semantic associations and structural characteristics. However, the presence of anomalous citation edges in practical applications (such as unrelated mutual citations, false citations, or fabricated self-citations) often indicates academic misconduct and may undermine the fairness, objectivity, and transparency of scientific research evaluation systems. Therefore, conducting graph link anomaly detection research in such citation networks can effectively identify potential illegal citation behaviors, and plays an important role in improving the credibility of scientific research data, improving academic supervision mechanisms, and purifying the academic ecological environment.
[0003] Currently, mainstream approaches for link anomaly detection in citation networks include statistical feature-based methods, graph embedding-based representation learning methods, and deep learning methods based on graph neural networks. Statistical methods typically rely on local structural features of edges (such as the number of common neighbors and the Jaccard coefficient) to score anomalies. While these methods offer advantages such as simplicity and computational efficiency, they have limited ability to model complex graph structures and are difficult to adapt to scenarios with rich attribute information or strong structural heterogeneity. Graph embedding methods attempt to map nodes or edges into a low-dimensional vector space and detect anomalies by calculating the anomaly of edges in this space. While these methods can capture global semantic relationships between nodes to a certain extent, they generally fail to simultaneously capture the deeper interactions between structure and attributes, and tend to overlook the impact of local perturbations on the overall graph structure. Deep learning methods based on graph neural networks, such as AANE, utilize GNNs to aggregate node features to learn richer structural representations. These methods then assess the degree of link anomaly by reconstructing edges or predicting their existence probabilities. Although this type of method has improved detection performance, it still has the following shortcomings: (1) Node features are prone to over-smoothing during multi-layer propagation, resulting in reduced differentiation between nodes and affecting the recognition of abnormal links; (2) Most methods are semi-supervised learning and are extremely dependent on label information, which greatly increases the cost; (3) In a network environment with a high degree of concealment of abnormal edges, existing models are prone to false detection or missed detection, and their practical application effect is limited. Summary of the Invention
[0004] In response to the above-mentioned deficiencies in the prior art, the academic paper citation anomaly detection method based on mutual information representation learning provided by the present invention solves the problem that traditional graph neural network methods are prone to excessive smoothing of node features during multi-layer propagation, thereby weakening the model's ability to distinguish between normal and abnormal links.
[0005] In order to achieve the above-mentioned object of the invention, the technical solution adopted by the present invention is: A method for detecting anomalies in academic paper citations based on mutual information representation learning is provided, which includes the following steps: S1. Obtain academic papers and construct their citation network, wherein the citation network includes an adjacency matrix consisting of citation relationships between academic papers and a node feature vector corresponding to each academic paper; S2. Input the adjacency matrix and all node feature vectors into the trained node embedding representation model to obtain the global embedding representation learned by maximizing mutual information; S3. Delete abnormal edges in the adjacency matrix to obtain a local adjacency matrix, and input it and the global embedding representation into the node embedding representation model to obtain the local embedding representation learned by maximizing mutual information; S4. Input the adjacency matrix and global embedding representation and the local adjacency matrix and local embedding representation into a trained anomaly detection model respectively to obtain two embedding representations, and concatenate the two to obtain an embedding vector; S5. Based on the embedding vector, calculate the scores of abnormal citations between academic paper pairs, sort them in descending order, and mark the paper pairs corresponding to the top k largest scores as abnormal citations.
[0006] Furthermore, the abnormal edge set in the adjacency matrix is deleted, and the method of reconstructing the adjacency matrix includes: K-means clustering is used to divide the embedded representation into a first preset number of categories to obtain the edge set between nodes of different pseudo-label categories , edge set The edge features in are expressed as: in, is the feature of edge e; and are the embedding representations of node i and node j respectively; is the edge set of the citation network; Use K-means clustering to cluster edge sets Perform clustering to obtain a second preset number of edge sub-sets; Set the edges Calculate the intersection with each sub-edge set respectively, select the maximum intersection as the abnormal edge set, delete the edges corresponding to the abnormal edge set in the adjacency matrix, and obtain the reconstructed adjacency matrix.
[0007] Furthermore, when maximizing mutual information to learn global embedding representation and local embedding representation, the expression of mutual information is: in, is the embedding representation of node i; is the neighbor set of node i; is mutual information; is the number of neighbors of node i; is the feature of the jth neighbor of node i; is the sigmoid function; To reconstruct the adjacency matrix; is the adjacency matrix The value of row i and column j in ; is the similarity between paper i and paper j; is the mean operation; is the node feature of negative sampling; is a soft-plus function; is the discriminator constructed by a neural network.
[0008] Furthermore, the expression of the embedding vector is: in, and These are the embedding representations of the two anomaly detection models respectively; is the splicing function.
[0009] Furthermore, based on the two embedding representations and the embedding vector, the preliminary reconstructed adjacency matrix is calculated , Abnormal Matrix And the final reconstructed adjacency matrix : in, is the activation function; T is the transpose; The loss function of the anomaly detection model is expressed as: in, is the hyperparameter used for balancing; U is the abnormal probability matrix; A is the adjacency matrix; is a norm; To reconstruct the adjacency loss; abnormal losses; is the Hadamard product; The expression of reconstruction adjacency loss is: in, and They are the loss values obtained from the reference relations existing in the original network and the loss values obtained from the reference relations that may exist in the original network; is a hyperparameter; The expression of abnormal loss is: Among them, c is the confidence parameter; It is a matrix summation operation.
[0010] Furthermore, based on the embedding vector, the expression for calculating the score of abnormal citations between academic paper pairs is: in, and They are The embedding vectors corresponding to academic papers u and v in [1]; Score for the number of unusual citations between academic papers u and v.
[0011] Furthermore, the node embedding representation model and the anomaly detection model are both graph convolutional neural networks; The graph convolutional neural network is a two-layer graph convolutional neural network. The convolution formula of each layer is: Where L is the Laplace matrix; Add the adjacency matrix of self-citations to the original citation network; I is the identity matrix; yes The diagonal matrix of is the activation function, and the input of the first convolutional layer is the initial attribute matrix , X is the matrix composed of all node feature vectors; A is the adjacency matrix; It is a one-layer graph convolutional network; is the learnable weight matrix of layer l; Embedding representation for the nodes in layer l.
[0012] Furthermore, the training method of the node embedding representation model includes: A1. Obtain a dataset of academic paper citations, including an adjacency matrix of scientific research paper citations and feature vectors generated by the semantic information of the papers. A2. Input the adjacency matrix and all eigenvectors from step A1 into the node embedding representation model. With the goal of maximizing mutual information, the node embedding representation model is trained using gradient descent and the Adam optimizer until the loss function converges. A3. Input the adjacency matrix and all eigenvectors from step A1 into the node embedding representation model converged in step A2 to obtain the global embedding representation learned by maximizing mutual information. A4. Delete the abnormal edges in the adjacency matrix in step A1 to obtain a local adjacency matrix, and input this matrix and the global embedding representation in step A3 into the node embedding representation model. With the goal of maximizing mutual information, use gradient descent and Adam optimizer to train the node embedding representation model converged in step A2 until the loss function converges. A5. Input the local adjacency matrix from step A4 and the global embedding representation from step A3 into the node embedding representation model converged in step A4 to obtain the local embedding representation learned by maximizing mutual information; The training methods for anomaly detection models include: The adjacency matrix in step A1, the global embedding representation in step A3, the local adjacency matrix in step A4, and the local embedding representation in step A5 are respectively input into an anomaly detection model. With the goal of minimizing the loss function, the anomaly detection model is trained using the gradient descent method and the Adam optimizer until the loss function converges.
[0013] The present invention's beneficial effects: This solution exploits multi-level structural information within the citation network from both a global and local structural perspective, and combines node attribute features to generate discriminative node embedding representations. This collaborative modeling from both a global and local perspective effectively mitigates the problem of over-smoothing node features during graph neural network training, improving the discriminability of the representation.
[0014] Compared to existing technologies, the anomaly detection method proposed in this paper can efficiently and accurately identify potentially anomalous citation behavior in research paper citation networks, without requiring any anomaly label information. This method is particularly well-suited for sparse network structures and can uncover citation relationships that may be anomalous or inconsistent with academic norms. This method is valuable in assisting the detection of academic misconduct, such as malicious cross-citations and citation manipulation.
[0015] Furthermore, this method, through a dual-perspective mutual information learning mechanism, effectively improves the expressive power of node representations and the sensitivity of identifying abnormal behavior, thereby enhancing the credibility and robustness of academic network structures. This method not only helps ensure the fairness of scientific research evaluation systems but also promotes authentic, high-quality academic collaboration and knowledge dissemination, possessing broad application prospects and practical significance in promoting the healthy development of scientific research ecosystems.
[0016] The present invention further explores the potential correlation between attribute features of papers, and adopts an unsupervised learning mechanism throughout the training of node embedding representation model and anomaly detection model to obtain node hidden representation that can characterize abnormal citation behavior; the entire training process does not rely on any abnormal label information, and node embedding learning and anomaly detection are both performed under an unsupervised framework, which significantly reduces the training cost of the model while still maintaining a high detection accuracy in anomaly detection tasks, and has good practical application value. BRIEF DESCRIPTION OF THE DRAWINGS
[0017] Figure 1 Flowchart of the academic paper citation anomaly detection method based on mutual information representation learning.
[0018] Figure 2 For different and The effect of the value on the experimental results.
[0019] Figure 3 Graph showing the impact of different components of the model on the experimental results. DETAILED DESCRIPTION
[0020] The specific embodiments of the present invention are described below to facilitate understanding of the present invention by those skilled in the art. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the appended claims, these changes are obvious, and all inventions and creations utilizing the concepts of the present invention are protected.
[0021] refer to Figure 1 , Figure 1 FIG. 1 shows a flowchart of an academic paper citation anomaly detection method based on mutual information representation learning; FIG. Figure 1 As shown, the method S includes steps S1 to S5.
[0022] In step S1, academic papers are obtained and their citation network is constructed. The citation network includes an adjacency matrix consisting of citation relationships between academic papers and a node feature vector corresponding to each academic paper. The citation network is represented as ,in Represents a paper node, that is, a node represents a paper, represents the reference edge set, and X represents the original feature representation of the node.
[0023] In step S2, the adjacency matrix and all node feature vectors are input into the trained node embedding representation model to obtain the global embedding representation learned by maximizing mutual information; During implementation, this scheme optimizes the mutual information to learn the global embedding representation and the local embedding representation. The mutual information expression is: in, is the embedding representation of node i; is the neighbor set of node i; is mutual information; is the number of neighbors of node i; is the feature of the jth neighbor of node i; is the sigmoid function; To reconstruct the adjacency matrix; is the adjacency matrix The value of row i and column j in ; is the similarity between paper i and paper j; is the mean operation; is the node feature of negative sampling; is a soft-plus function; is the discriminator constructed by a neural network.
[0024] In step S3, the abnormal edges in the adjacency matrix are deleted to obtain the local adjacency matrix, and the local adjacency matrix is input into the node embedding representation model together with the global embedding representation to obtain the local embedding representation learned by maximizing the mutual information; During implementation, this solution preferably deletes abnormal edge sets in the adjacency matrix, and the method for reconstructing the adjacency matrix includes: K-means clustering is used to divide the embedded representation into a first preset number of categories to obtain the edge set between nodes of different pseudo-label categories , edge set The edge features in are expressed as: in, is the feature of edge e; and are the embedding representations of node i and node j respectively; is the edge set of the citation network; Use K-means clustering to cluster edge sets Perform clustering to obtain a second preset number of edge sub-sets; Set the edges Calculate the intersection with each sub-edge set respectively, select the maximum intersection as the abnormal edge set, delete the edges corresponding to the abnormal edge set in the adjacency matrix, and obtain the reconstructed adjacency matrix.
[0025] The first preset number is equal to the second preset number, and both are 2.
[0026] This solution implements a dual-perspective embedding mechanism through steps S2 and S3, performing embedding learning from both a local and a global perspective. The local perspective focuses on the local structural information of node neighbors, preserving the semantic features of nodes in the local topology, while the global perspective captures the consistency of global dependencies between nodes based on the overall topological structure of the graph. By introducing the mutual information maximization constraint, the discriminability of the embedded representations between different node classes is enhanced, effectively alleviating the problem of decreased discriminative ability caused by oversmoothing during feature propagation.
[0027] In step S4, the adjacency matrix and global embedding representation, and the local adjacency matrix and local embedding representation are respectively input into a trained anomaly detection model to obtain two embedding representations, which are then concatenated to obtain an embedding vector; the expression of the embedding vector is: in, and These are the embedding representations of the two anomaly detection models respectively; is the splicing function.
[0028] In step S5, the score of abnormal citations between pairs of academic papers is calculated based on the embedding vector: in, and They are The embedding vectors corresponding to academic papers u and v in [1]; Score for the number of unusual citations between academic papers u and v.
[0029] Sort all the scores in descending order, and mark the paper pairs corresponding to the top k largest scores as abnormal citations; the value of k can be adjusted according to different network sizes.
[0030] In one embodiment of the present invention, the preliminary reconstructed adjacency matrix is calculated based on the two embedding representations and the embedding vector. , Abnormal Matrix And the final reconstructed adjacency matrix : in, is the activation function; T is the transpose; The loss function of the anomaly detection model is expressed as: in, is the hyperparameter used for balancing; U is the abnormal probability matrix; A is the adjacency matrix; is a norm; To reconstruct the adjacency loss; abnormal losses; is the Hadamard product; since this task mainly considers abnormal edges, we use and Construct regularization terms to prevent overfitting.
[0031] The expression of reconstruction adjacency loss is: in, and They are the loss values obtained from the reference relations existing in the original network and the loss values obtained from the reference relations that may exist in the original network; is a hyperparameter used to set the proportion of citation relations existing in the original structure in the reconstruction loss.
[0032] The expression of abnormal loss is: Among them, c is the confidence parameter; It is a matrix summation operation.
[0033] In this solution, for abnormal losses , which is used to measure the difference between the similarity between papers and the citation anomaly relationship. From the perspective of the task goal, it is worth noting that this scheme only focuses on the edges that exist in the citation relationship. At the same time, the citation relationship is mostly sparse, and the edges without citations account for the majority, while the outliers of the edges with citations only account for a very small part of the entire similarity matrix. Therefore Only reference relations existing in the original graph structure are considered.
[0034] The anomaly probability learned at the beginning is not convincing enough because the network has not fully converged. Therefore, a confidence parameter c is introduced to represent the confidence of the anomaly probability in each learning iteration. Relationship with the original reference The confidence parameter is calculated by the difference between the two, that is, the closer c is to 1, the closer the node similarity is to the adjacency matrix. The smaller the difference, the better the fit.
[0035] During implementation, this solution preferably uses graph convolutional neural networks as both the node embedding representation model and the anomaly detection model; The graph convolutional neural network is a two-layer graph convolutional neural network. The convolution formula of each layer is: Where L is the Laplace matrix; Add the adjacency matrix of self-citations to the original citation network; I is the identity matrix; yes The diagonal matrix of is the activation function, and the input of the first convolutional layer is the initial attribute matrix , X is the matrix composed of all node feature vectors; A is the adjacency matrix; It is a one-layer graph convolutional network; is the learnable weight matrix of layer l; Embedding representation for the nodes in layer l.
[0036] In one embodiment of the present invention, the training method of the node embedding representation model includes: A1. Obtain an academic paper citation dataset, which includes an adjacency matrix composed of scientific research paper citations and feature vectors generated by the paper's semantic information.
[0037] In this embodiment, the scientific research paper citation datasets used are publicly available. The two academic citation datasets: Cora contains 2708 papers and 10556 citations, and Citeseer contains 3327 papers and 9104 citations. The specific information of the datasets is shown in Table 1.
[0038] Table 1 Statistics of scientific research citation network data A2. Input the adjacency matrix and all eigenvectors from step A1 into the node embedding representation model. With the goal of maximizing mutual information, set the learning rate and Weight_decay to 5e-4, and train the node embedding representation model using gradient descent and the Adam optimizer until the loss function converges.
[0039] A3. Input the adjacency matrix and all eigenvectors from step A1 into the node embedding representation model converged in step A2 to obtain the global embedding representation learned by maximizing mutual information. The node embedding representation model is a graph convolutional neural network, which can encode the attribute network by smoothing and propagating the attribute information on the network and the nonlinear transformation of the input attributes. The graph convolutional neural network is a two-layer graph convolutional neural network. The formula for each layer is as follows: In this example, Select the LeakyReLU activation function, for the Cora dataset, the input , , Represents the identity matrix. Since a two-layer graph convolutional neural network is used, the corresponding two parameter matrices , , the final node embedding vector matrix ; For the Citeseer dataset, the input , , since a two-layer graph convolutional neural network is used, the corresponding two parameter matrices , , the final node embedding vector matrix .
[0040] A4. Delete the abnormal edges in the adjacency matrix in step A1 to obtain the local adjacency matrix. The specific edge deletion information of the two datasets is shown in Table 2.
[0041] Table 2 Statistics of abnormal edge deletion accuracy The local adjacency matrix and the global embedding representation in step A3 are input into the node embedding representation model. With the goal of maximizing mutual information, the learning rate and Weight_decay are both set to 5e-4. The node embedding representation model converged in step A2 is trained using the gradient descent method and the Adam optimizer until the loss function converges.
[0042] A5. Input the local adjacency matrix from step A4 and the global embedding representation from step A3 into the node embedding representation model converged in step A4 to obtain the local embedding representation learned by maximizing mutual information; The training methods for anomaly detection models include: The adjacency matrix in step A1, the global embedding representation in step A3, the local adjacency matrix in step A4, and the local embedding representation in step A5 are respectively input into an anomaly detection model. With the goal of minimizing the loss function, the learning rate and Weight_decay are both set to 5e-4. The anomaly detection model is trained using gradient descent and the Adam optimizer until the loss function converges.
[0043] In order to evaluate the academic paper citation anomaly detection method of the present invention, it is compared with the most advanced methods in the field of connection anomaly detection, including AANE, DGI, GMI, GMI++ and RGSE. The GMI++ method is a further version of GMI, which introduces the PPMI matrix in the loss design. In order to ensure the statistical credibility of the results, the average accuracy (%) and recall (%) of 5 runs are selected as the indicators for measuring the anomaly detection task. At the same time, since the method proposed in the present invention is an unsupervised method, the anomaly labels are only used for evaluation in the test phase. In order to test the effectiveness of the method, the common Precision index, where k can be set to 50, 100 or 200. In addition, this embodiment has two variants of anomaly detection. If all heterogeneous edges are deleted at once, it is called New; if an iterative strategy is used to gradually correct the graph structure, for example, it is iterated N times, and each time it is deleted from the graph Heterogeneous edges, until a total of This iterative edge removal strategy is denoted as New-D. In the experiment, the number of iterations N is set to 10. The experimental results of the Cora dataset are shown in Table 3: Table 3. Cora dataset reference anomaly detection results The experimental results of the Citeseer dataset are shown in Table 4: Table 4 Citeseer dataset citation anomaly detection results Since the number of anomalous edges in a citation network is not necessarily an integer when divided by the number of training iterations, the recall of the "New-D" method cannot be compared in the evaluation metrics. Therefore, the recall metric in Tables 3 and 4 is represented by a " / ". The results in Tables 3 and 4 show that the proposed method achieves optimal performance in both key performance metrics—accuracy (columns 2-4 in Tables 3 and 4) and recall. Furthermore, the proposed method demonstrates superior performance under different precision thresholds for accuracy, fully demonstrating its robustness and detection capabilities in the task of anomalous citation detection.
[0044] Figure 2 The two parameters in the loss function are discussed to reconstruct the loss coefficient and abnormal loss coefficient Impact on model accuracy. This solution will and The values of are set to 0.1, 0.3, 0.5, 0.7 and 0.9 respectively, and the accuracy changes on the two datasets are recorded. The experimental results show that in the citation network datasets (Cora, CiteSeer), the model has a good effect on the reconstruction loss coefficient. More sensitive. Specifically, on the Cora dataset, as Increasing from 0.1 to 0.9, the accuracy gradually increased from 36% to 50% (a significant increase of 14%), while the abnormal loss coefficient The impact on performance is relatively limited, with fluctuations of less than 2%. In contrast, the CiteSeer dataset shows the opposite trend - when When the value of α increases, the accuracy drops from 42% to 34% (a decrease of 8%), but The change of still only causes fluctuations of less than 3%. This shows that in the citation network anomaly detection task, the proportion of reconstruction loss term (i.e. The setting of ) has a great impact on the overall performance of the model. Reasonable configuration of this parameter can help improve detection accuracy.
[0045] In order to systematically analyze the contribution of each component in anomaly detection methods, a comprehensive ablation study is conducted, e.g. Figure 3As shown. Three variant models are set up: (1) no-gl (does not include global view mutual information maximization), this variant removes the mutual information maximization process under the global perspective and performs clustering directly on the normalized feature matrix; (2) no-lo (does not include local view mutual information maximization), this variant removes the clustering operator and the local view mutual information learning mechanism, and maintains the global perspective modeling; (3) no-op (does not include negative sampling optimization), adopts the subgraph-level negative sampling strategy of the global view, which randomly selects non-adjacent nodes based on neighborhood similarity. The complete model is a standard version containing all the above components, which is used to compare the performance with each variant. To ensure fairness of the comparison, all models are trained under the same experimental settings and hyperparameter conditions, and the average accuracy of the five experiments obtained on the test set is finally used as the evaluation indicator.
[0046] like Figure 3 As shown in the figure, the complete model proposed in this paper demonstrates significant and stable performance advantages on both benchmark datasets. Furthermore, comparative experiments were conducted by removing each key component from the model one by one. The results showed that the model accuracy decreased to varying degrees, verifying that each module played an indispensable role in the overall performance. Therefore, it can be concluded that the model configuration under the complete structure can achieve optimal performance and has good structural rationality and synergistic effectiveness.
[0047] In summary, the academic paper citation anomaly detection method proposed in the present invention effectively enhances the discriminative ability of node representation by introducing a mutual information maximization mechanism, thereby achieving efficient identification of abnormal citation behavior under an unsupervised learning framework. Experimental results show that the proposed method achieves optimal performance in key indicators such as accuracy and recall rate, verifying its strong robustness and practicality in complex citation network structures. The method is applicable to large-scale academic citation network scenarios in actual scientific research environments, and can automatically identify phenomena such as mutual citation relationships that lack real academic connections, false self-citation behaviors, and forged citations. In some cases, such abnormal citation behaviors may be associated with serious problems such as academic misconduct and paper fraud, thereby affecting the fairness and data transparency of the scientific research evaluation system.
[0048] Therefore, in the task of detecting abnormal links in citation networks, the detection method provided by the present invention not only helps to timely discover potential abnormal citation behaviors and academic irregularities, but also helps to improve the overall credibility of scientific research data resources and promote the healthy development of the academic ecological environment. It has good practical application value and social benefits.
Claims
1. A method for detecting anomaly in academic paper citations based on mutual information representation learning, characterized by: Including steps: S1. Obtain academic papers and construct their citation network, wherein the citation network includes an adjacency matrix consisting of citation relationships between academic papers and a node feature vector corresponding to each academic paper; S2. Input the adjacency matrix and all node feature vectors into the trained node embedding representation model to obtain the global embedding representation learned by maximizing mutual information; S3. Delete abnormal edges in the adjacency matrix to obtain a local adjacency matrix, and input it and the global embedding representation into the node embedding representation model to obtain the local embedding representation learned by maximizing mutual information; S4. Input the adjacency matrix and global embedding representation and the local adjacency matrix and local embedding representation into a trained anomaly detection model respectively to obtain two embedding representations, and concatenate the two to obtain an embedding vector; S5. Based on the embedding vector, calculate the scores of abnormal citations between academic paper pairs, sort them in descending order, and mark the paper pairs corresponding to the top k largest scores as abnormal citations.
2. The method for detecting anomaly in academic paper citations according to claim 1, characterized in that: Methods for deleting abnormal edge sets in the adjacency matrix and reconstructing the adjacency matrix include: K-means clustering is used to divide the embedded representation into a first preset number of categories to obtain the edge set between nodes of different pseudo-label categories , edge set The edge features in are expressed as: in, is the feature of edge e; and are the embedding representations of node i and node j respectively; is the edge set of the citation network; Use K-means clustering to cluster edge sets Perform clustering to obtain a second preset number of edge sub-sets; Set the edges Calculate the intersection with each sub-edge set respectively, select the maximum intersection as the abnormal edge set, delete the edges corresponding to the abnormal edge set in the adjacency matrix, and obtain the reconstructed adjacency matrix.
3. The method for detecting anomaly in academic paper citations according to claim 1, characterized in that: When maximizing mutual information to learn global embedding representation and local embedding representation, the expression of mutual information is: in, is the embedding representation of node i; is the neighbor set of node i; is mutual information; is the number of neighbors of node i; is the feature of the jth neighbor of node i; is the sigmoid function; To reconstruct the adjacency matrix; is the adjacency matrix The value of row i and column j in ; for exist The proportion in is calculated by the similarity between paper i and paper j; is the mean operation; is the node feature of negative sampling; is a soft-plus function; is the discriminator constructed by a neural network.
4. The method for detecting anomaly in academic paper citations according to claim 1, characterized in that: The expression of the embedding vector is: in, and These are the embedding representations of the two anomaly detection models respectively; is the splicing function.
5. The method for detecting anomaly in academic paper citations according to claim 4, characterized in that: Based on the two embedding representations and the embedding vector, calculate the preliminary reconstruction adjacency matrix , Abnormal Matrix And the final reconstructed adjacency matrix : in, is the activation function; T is the transpose; The loss function of the anomaly detection model is expressed as: in, is the hyperparameter used for balancing; U is the abnormal probability matrix; A is the adjacency matrix; is the 1-norm; To reconstruct the adjacency loss; abnormal losses; is the Hadamard product; The expression of reconstruction adjacency loss is: in, and They are the loss values obtained from the reference relations existing in the original network and the loss values obtained from the reference relations that may exist in the original network; is a hyperparameter; The expression of abnormal loss is: Among them, c is the confidence parameter; It is a matrix summation operation.
6. The method for detecting anomaly in academic paper citations according to claim 4, characterized in that: Based on the embedding vector, the expression for calculating the score of abnormal citations between academic paper pairs is: in, and They are The embedding vectors corresponding to academic papers u and v in [1]; Score for the number of unusual citations between academic papers u and v.
7. The method for detecting anomaly in academic paper citations according to any one of claims 1 to 6, characterized in that: The node embedding representation model and the anomaly detection model are both graph convolutional neural networks; The graph convolutional neural network is a two-layer graph convolutional neural network. The convolution formula of each layer is: Where L is the Laplace matrix; Add the adjacency matrix of self-citations to the original citation network; I is the identity matrix; yes The diagonal matrix of is the activation function, and the input of the first convolutional layer is the initial attribute matrix , X is the matrix composed of all node feature vectors; A is the adjacency matrix; It is a one-layer graph convolutional network; is the learnable weight matrix of layer l; Embedding representation for the nodes in layer l.
8. The method for detecting anomaly in academic paper citations according to any one of claims 1 to 6, characterized in that: The training method of the node embedding representation model includes: A1. Obtain a dataset of academic paper citations, including an adjacency matrix of scientific research paper citations and feature vectors generated by the semantic information of the papers. A2. Input the adjacency matrix and all eigenvectors from step A1 into the node embedding representation model. With the goal of maximizing mutual information, the node embedding representation model is trained using gradient descent and the Adam optimizer until the loss function converges. A3. Input the adjacency matrix and all eigenvectors from step A1 into the node embedding representation model converged in step A2 to obtain the global embedding representation learned by maximizing mutual information. A4. Delete the abnormal edges in the adjacency matrix in step A1 to obtain a local adjacency matrix, and input this matrix and the global embedding representation in step A3 into the node embedding representation model. With the goal of maximizing mutual information, use gradient descent and Adam optimizer to train the node embedding representation model converged in step A2 until the loss function converges. A5. Input the local adjacency matrix from step A4 and the global embedding representation from step A3 into the node embedding representation model converged in step A4 to obtain the local embedding representation learned by maximizing mutual information; The training methods for anomaly detection models include: The adjacency matrix in step A1, the global embedding representation in step A3, the local adjacency matrix in step A4, and the local embedding representation in step A5 are respectively input into an anomaly detection model. With the goal of minimizing the loss function, the anomaly detection model is trained using the gradient descent method and the Adam optimizer until the loss function converges.