Graph data preprocessing method, system and equipment and medium

Abstract

The invention provides a graph data preprocessing method, system and device and a medium, and belongs to the technical field of graph data processing.The method comprises the following steps that a plurality of disturbance edges of graph data to be processed are removed, and a rough graph is obtained; randomly selecting some edges from the plurality of disturbance edges, and adding the selected edges into the rough image to generate a to-be-predicted image; using a depth map mutual information model (DGI) to learn mutual information between the rough image and the representation of the rough image and mutual information between the image to be predicted and the representation of the image to be predicted; whether the difference value of the two pieces of mutual information is smaller than a threshold value or not is judged, if not, multiple edges are selected from the remaining disturbance edges to be added into the to-be-predicted image to generate a new to-be-predicted image, whether the difference value of the new to-be-predicted image and the mutual information represented by the new to-be-predicted image and the mutual information between the rough image and the rough image is smaller than the set threshold value or not is judged again, and if yes, the new to-be-predicted image is generated. And obtaining a clean image. According to the method, tiny disturbance structures caused by adversarial attacks in the graph data can be removed, and clean graph data is obtained and can be conveniently applied to downstream tasks.

Description

Technical Field

[0001] This invention belongs to the field of graph representation learning technology, specifically relating to a graph data preprocessing method, system, device, and medium. Background Technology

[0002] Graph data is a ubiquitous data structure capable of representing various entities and their complex interconnections. Graph data can reflect various real-world networks, including protein networks, transportation networks, and social networks. Currently, Graph Neural Networks (GNNs) have become a powerful tool for effectively analyzing graphs due to their remarkable performance in tasks such as node classification, link prediction, and graph classification. Especially after introducing unsupervised learning into GNNs, learning can be performed without relying on labels, while maintaining almost no change in the representation learning quality of GNNs.

[0003] However, many studies indicate that graph neural networks (GNNs) exhibit significant vulnerability when subjected to malicious adversarial attacks. In other words, even minor perturbations (polluted graph data) in the structure of the graph data input into a GNN can lead to incorrect predictions. For example, in citation networks, nodes represent papers, and edges represent citation relationships. Tasks typically include paper topic classification, research impact prediction, or knowledge dissemination modeling. If an attacker can perturb the citation network (e.g., forging cross-domain citation relationships or deleting key citation edges), it could have multiple impacts: firstly, it could directly lead to incorrect node classification results, distorting research topic segmentation or knowledge structure analysis; secondly, it could mislead academic recommendation and retrieval systems, causing low-quality or even fraudulent documents to be incorrectly recommended, thus amplifying information pollution; furthermore, in real-world applications, such networks are often used as foundational data for scientometrics and research evaluation, and attacks on them could lead to biases in academic evaluation or resource allocation.

[0004] Currently, existing technologies employ similarity-based detection methods to determine whether a node has been perturbed, which involves removing edges between nodes whose representations have very low similarity. However, these methods perform poorly in defending against adversarial attacks. This is because adversarial attacks are more likely to connect nodes that are far apart, and perturbed edges may cause the model to learn similar representations for nodes that were not originally adjacent, thus failing to effectively identify subtle perturbation structures and resulting in poor preprocessing performance of contaminated citation networks. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies in removing minute perturbations in citation network graph data, this invention provides a graph data preprocessing method, comprising the following steps: Remove multiple perturbation edges from the graph data to be processed to obtain a rough map; Randomly select some edges from multiple perturbation edges and add them to the coarse map to generate the graph to be predicted; use the Depth Graph Mutual Information (DGI) model to learn the mutual information between the coarse map and its representation, as well as the mutual information between the graph to be predicted and its representation; use graph contamination to determine whether the difference between the two mutual information values ​​is less than a threshold. If not, select multiple edges from the remaining perturbation edges and add them to the graph to be predicted to generate a new graph to be predicted. Then, determine again whether the difference between the mutual information of the new graph to be predicted and its representation and the mutual information between the coarse map and its representation is less than a set threshold. If so, a clean graph is obtained.

[0006] Preferably, the step of removing multiple perturbation edges from the citation network dataset to obtain a coarse map includes the following steps: Randomly select a node from the citation network dataset to be processed. i The Restart Random Walk (RWR) algorithm is used to process the nodes. i A random walk is performed, and the nodes and edges of the walk are combined to form positive samples. The collected positive samples are represented as subgraphs. The positive sample subgraphs are groups of similar papers with the same theme and their citation links. Negative sample subgraphs are obtained by randomly selecting positive sample subgraphs from nodes other than the positive samples. The negative sample subgraphs are groups of papers with completely different fields and their abnormal citation links. Scoring the positive and negative sample subplots separately includes the following steps: Learning nodes using the GCN model i The overall embedding representation of the subgraph; Using a deep neural network (DNN) to connect nodes i Mapping to the same embedding space as the overall embedding representation of the subgraph yields the nodes. i Embedded representation; Based on the overall embedding representation of the subgraph and nodes i The embedded representation is used, and the scores of the graph data nodes are obtained by combining the bilinear scoring function; The negative of the difference between the positive sample score and the negative sample score is taken as the outlier score; based on the graph data nodes. i The abnormal score derivation graph data is used to derive the abnormal scores of the edges, sort the abnormal scores of the edges, and remove the top ones. k By analyzing the edges, we can obtain a rough map.

[0007] Preferably, the abnormal scores of the edges of the graph data are derived from the abnormal scores of the graph data nodes using the following formula: ; In the formula, S i For nodes i Abnormal scores, S j For nodesj Abnormal scores, A i，j =1 indicates a node i and nodes j There are edges between them. A i，j =0 indicates a node i and nodes j There is no edge between them.

[0008] Preferably, the learning nodes are adopted using the GCN model. i The subgraph embedding representation is performed using the following formula: ; In the formula, k The subgraph represents the first k 1 node K Indicates the number of nodes in the subgraph. H k Indicates the first k The representation of each node.

[0009] Preferably, the node i The embedded representation is obtained by the following formula: ; In the formula, h i (l-1) Represents a node i In the l- The representation of layer 1, express l Parameters of layer -1 It is the ReLU activation function. h i Table Node i The overall embedding representation of the subgraph.

[0010] Preferably, after obtaining the rough map, the difference between the two mutual information values ​​is determined using graph contamination degree, whereby the graph contamination degree is: ; In the formula, G This is a rough sketch. G 'This is the image to be predicted.' I ( G , f θ ( G ))for G and f θ ( G Mutual information between them f θ ( G () is a rough representation.I ( G ', f θ ( G '))for G 'and f θ ( G Mutual information between ') f θ ( G ') represents the graph to be predicted, GPD( i) This represents the difference in mutual information between the rough map and the map to be predicted.

[0011] This invention also provides a method for progressively restoring coarse map data to clean map data, comprising: The rough map acquisition module is used to remove multiple perturbation edges from the graph data to be processed to obtain a rough map; The graph cleanup module uses perturbation edges as candidate edges, randomly selects some edges from multiple candidate edges and adds them to the coarse map to generate the graph to be predicted; it uses the Deep Graph Mutual Information (DGI) model to learn the mutual information between the coarse map and the representation obtained from the coarse map, as well as the mutual information between the graph to be predicted and the representation obtained from the graph to be predicted; it uses graph contamination degree to determine whether the difference between the two mutual information values ​​is less than a threshold. If not, it continues to select multiple edges from the candidate edges and add them to the graph to be predicted to generate a new graph to be predicted, and again determines whether the new difference is less than a set threshold. If so, it updates the coarse map until a clean citation network is obtained.

[0012] The present invention also provides a computer device including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to perform the graph data preprocessing method.

[0013] The present invention also provides a computer-readable storage medium storing a computer program adapted for loading by a processor to execute the graph data preprocessing method.

[0014] The graph data preprocessing method provided by this invention has the following beneficial effects: This invention obtains a coarse graph by removing most of the disturbed edges in the citation network; it generates a graph to be predicted by repeatedly adding perturbation edges to the coarse graph; it obtains the mutual information between the coarse graph and its representation, as well as the mutual information between the graph to be predicted and its representation, through DGI; it cleans the citation network by judging the difference between the two mutual information values ​​each time a perturbation edge is added; and through the iterative cleansing process, it gradually restores accurate node representations, removes minor perturbation structures from the graph data, and finally obtains clean graph data. This not only improves the accuracy of node classification but also ensures more reliable results for literature retrieval and recommendation, maintaining the authenticity and credibility of knowledge dissemination and academic evaluation. Attached Figure Description

[0015] To more clearly illustrate the embodiments and design schemes of the present invention, the accompanying drawings required for this embodiment will be briefly described below. The drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 This is a flowchart of a graph data preprocessing method according to an embodiment of the present invention. Detailed Implementation

[0017] To enable those skilled in the art to better understand and implement the technical solutions of the present invention, the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0018] In the description of this invention, it should be understood that the terms "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," "outer," "axial," "radial," and "circumferential" indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing the technical solution of this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0019] Furthermore, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. In the description of this invention, it should be noted that, unless otherwise explicitly specified or limited, the terms "connected" or "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances. In the description of this invention, unless otherwise stated, "a plurality of" means two or more, which will not be elaborated further here.

[0020] Example This invention provides a graph data preprocessing method, specifically as follows: Figure 1 As shown, it includes the following steps: Step 1: Obtain the overall embedding representation of the subgraphs and nodes of the graph data. i Embedded representation, based on the global embedding representation of subgraphs and nodes i Embedded representation of obtaining graph data nodes i Abnormal scores. The specific steps include the following:

[0021] S1: Randomly select a node from the graph data to be processed. i The Restart Random Walk (RWR) algorithm is used to process the nodes. i A random walk is performed, and the nodes and edges of the walk are combined to form positive samples. The collected positive samples are represented as subgraphs, which consist of groups of similar papers with the same topic and their citation links. However, learning using only positive sample pairs may lead to model collapse. Therefore, negative sample subgraphs are obtained by randomly selecting positive sample subgraphs from nodes other than the positive samples. These negative sample subgraphs consist of groups of papers with completely different fields and their anomalous citation links.

[0022] S2: Learning nodes using the GCN model i The subgraph embedding representation is as follows: First, this invention requires learning subgraphs. G The embedding representation, since the Graph Convolutional Network (GCN) can aggregate information from nodes in a local subgraph and transform high-dimensional attributes into a low-dimensional embedding space, this invention employs an efficient GCN model as the model for learning node embeddings. G The embedding formula can be expressed as: ; In the formula, It is the ReLU activation function. Di Represents a node i The degree, A i Represents a node i The adjacency matrix of the subgraph. l Indicates the number of floors. express l -1 floor i Node representation, express l Parameters of layer -1.

[0023] Next, this invention applies an average pooling function as the readout function to obtain the subgraph. G Overall Embedded Representation .

[0024] ; In the formula, k The subgraph represents the first k 1 node K Indicates the number of nodes in the subgraph. H k Indicates the first k The representation of each node.

[0025] S3: Use a deep neural network (DNN) to connect nodes i Mapping to the same embedding space as the overall embedding representation of the subgraph yields the nodes. i Embedded representation.

[0026] This invention uses deep neural networks (DNNs) to compute nodes. i The network will have nodes. i Mapping to The same embedding space, specifically, nodes i The embedding representation is shown below: ; In the formula, h i (l-1) Represents a node i In the l- The representation of layer 1, express l Parameters of layer -1 It is the ReLU activation function. h i Table Node i The overall embedding representation of the subgraph.

[0027] S4: Based on the positive sample subgraph Overall embedding representation and nodesi The embedding representation is used, and the positive sample score of the graph data node is obtained by combining it with the bilinear scoring function. The specific formula is as follows: ; In the formula, This represents the sigmoid activation function. Subgraph The overall embedded representation, W Represents a learnable matrix. h i Represents a node i Embedded representation, It should be close to 1.

[0028] Learning using only positive sample pairs may cause the model to collapse; therefore, negative sample scores are introduced. Its negative sample subgraph It is obtained by randomly selecting positive sample subgraphs from other nodes; for negative sample pairs... The specific formula for negative sample scores that are close to 0 is as follows: ; In the formula, j Indicates except i Any random node outside of the range.

[0029] Finally, the node i The outlier score is defined as the negative of the difference between the positive sample score and the negative sample score: ; Abnormal scores The higher the value, the more likely the node is to be active. i The more likely it is to be an abnormal node.

[0030] Step 2: Based on the abnormal scores of node i in the graph data, deduce the abnormal scores of the edges in the graph data, sort the abnormal scores of the edges, and remove the first k edges to obtain a rough graph.

[0031] This invention first requires obtaining a method to eliminate most perturbation edges, and then refines it step by step. Adversarial attacks typically connect nodes that are far apart to maximally disrupt the homogeneity of graph data, usually resulting in high anomaly scores for the two connected endpoints (nodes). Therefore, if the detected endpoints of an edge have high anomaly scores, it indicates that the edge is highly likely to be perturbed. In practice, this invention relies on the anomaly scores of endpoints to derive the anomaly matrix of edges, thereby obtaining the anomaly scores of the edges. The specific anomaly matrix is ​​shown below:

[0032] ; In the formula, S i For nodesi Abnormal scores, S j For nodes j Abnormal scores, A i，j =1 indicates a node i and nodes j There are edges between them. A i，j =0 indicates a node i and nodes j There is no edge between them.

[0033] Next, sort the edges in descending order of abnormal scores, and then remove the previous ones. k Find the most unusual edge to generate a rough map. G .

[0034] Step 3: Randomly select some edges from multiple perturbation edges and add them to the coarse map to generate the graph to be predicted; use the Deep Graph Infomax (DGI) model to learn the mutual information between the coarse map and its representation, as well as the mutual information between the graph to be predicted and its representation; use graph contamination to determine whether the difference between the two mutual information values ​​is less than a threshold. If not, select multiple edges from the remaining perturbation edges and add them to the graph to be predicted to generate a new graph to be predicted (a gradually cleaned coarse map), and again determine whether the difference between the mutual information of the new graph to be predicted and its representation and the mutual information between the coarse map and its representation is less than a set threshold. If so, a clean graph is obtained.

[0035] This invention proposes determining edge perturbation by measuring changes in mutual information. Intuitively, the goal of adversarial attacks is to disrupt the representation, thereby reducing the performance of downstream tasks. This demonstrates that the representation learned from the graph to be predicted cannot adequately reflect a graph containing adversarial edges. Therefore, this invention proposes graph contamination as a measure of the degree of contamination in graph data. The formula for graph contamination is as follows:

[0036] ; In the formula, G This is a rough sketch. G 'This is the image to be predicted.' I ( G , f θ ( G ))for G and f θ ( G Mutual information between them f θ ( G () is a rough representation. I ( G ',f θ ( G '))for G 'and f θ ( G Mutual information between ') f θ ( G ') represents the graph to be predicted, GPD( i) This represents the difference in mutual information between the rough map and the map to be predicted.

[0037] lower GPD ( i The value of ) indicates that the examined graph data contains perturbation edges with a higher risk, according to GPD ( i The value of ) allows for continuous fine-tuning and purification of the rough map, enabling the assessment of the map data during the purification process. The degree of contamination is determined by whether the number of true edges in the purified graph data reaches a set threshold, thus obtaining the target clean graph.

[0038] To verify the performance of the method for processing graph data, experiments were first conducted on three benchmark datasets: Cora, Citeseer, and PubMed. These datasets are citation network datasets, where nodes represent documents and edges represent citation links between them. In Cora and Citeseer, features are binary indicators reflecting the presence of specific words in each document. In PubMed, features are based on TF-IDF (Term Frequency-Inverse Document Frequency) values, capturing the importance of terms in documents. The edges in these datasets were created based on citation relationships, where a directed edge from one node to another indicates that the first document cites the second document. Detailed feature information for the datasets is shown in Table 1.

[0039] Secondly, in order to evaluate the effectiveness of the method of the present invention, the present invention will use the following seven benchmark methods and the method of the present invention to clean the graph data in the above three datasets. The seven benchmark methods are as follows: Projected Gradient Descent (PGD): This method uses the principle of maximizing mutual information to learn unsupervised node representations.

[0040] JACCARD Similarity Method: Removes edges with high distances based on the JACCARD similarity of the representation.

[0041] Dominant (Deep Anomaly Detection on Attributed Networks): This method uses a deep autoencoder to detect perturbed edges.

[0042] CoLA (Contrastive self-supervised Learning framework for anomaly detection on attributed networks) learns anomalies by capturing the relationships between each node and its neighborhood substructure in an unsupervised manner, thereby identifying perturbation edges.

[0043] ANEMONE (ANomaly detection framework with Multi-scale contrastive learning): This paper proposes a contrastive learning-based method for learning anomaly nodes.

[0044] Sub-CR (Subgraph-based multi-view self-supervised framework which contains a Contrastive-based module and a Reconstruction-based module): proposes a dual-view contrastive learning module.

[0045] GRADATE (GRaph Anomaly Detection framework via multi-scale contrastive learning networks with newly added subgraph-subgraph contrast and Augmented view): A framework for graph anomaly detection using multi-scale contrastive learning networks, which includes newly added subgraph-subgraph contrast and augmented view, detecting anomalies through subgraph-subgraph contrast, subgraph-node contrast, and node-node contrast.

[0046] Next, the seven benchmark methods mentioned above and the method of this invention are used to clean the graph data in the three datasets. Then, the cleaned graph data is used to train PGD for attack (PGD is an attack method, not a defense method. It applies convex relaxation optimization to perform gradient-based adversarial attacks on discrete graph data. The attack principle includes calculating the gradient of the input data and adding perturbations in the gradient direction. Through multiple iterations, adversarial examples are gradually constructed, and then training representation is performed on the perturbated graph structure). The performance of the cleaned graph data is judged based on the training effect. The results are shown below, see Table 2 for details.

[0047] (1) Regardless of the perturbation ratio, the method of the present invention is always superior to the seven benchmark methods. This is because the data processed by the seven benchmark methods is equivalent to the rough map obtained by the present invention, while the method of the present invention further identifies structures containing more “real edges” based on the rough map, thereby achieving better classification results than the seven benchmark methods.

[0048] (2) PGD performs worse than other algorithms in most cases. This indicates that the representation learned by PGD lacks robustness, highlighting the necessity and importance of exploring other algorithms to learn representations with better robust performance.

[0049] (3) The JACCARD and Dominant methods also perform poorly in defending against PGD attacks. This is mainly because these methods focus primarily on structural errors rather than anomalous patterns during training, making them more sensitive to the effects of perturbation edges. In contrast, the method of this invention avoids directly relying on structural information to determine whether an edge is anomalous, thus performing better in defending against attacks.

[0050] (4) As the perturbation increases, the performance of the method of the present invention degrades much more slowly on all datasets, indicating that the method of first selecting a coarse map and then fine-tuning the model based on different maps is less sensitive to perturbation.

[0051] (5) CoLA, Sub-CR, and GRADATE are three relatively new algorithms that use graph contrastive learning to detect graph anomalies. These methods have shown strong robustness on most datasets. However, they face a common problem: the obtained graph still contains some "spurious" edges, which may interfere with representation learning, leading to inaccurate representations and thus affecting the performance of downstream tasks.

[0052] This invention also provides a graph data preprocessing system, comprising: The rough map acquisition module is used to remove multiple perturbation edges from the graph data to be processed to obtain a rough map; The graph cleanup module generates a graph to be predicted by randomly selecting some edges from multiple perturbation edges and adding them to a coarse map; it learns the mutual information between the coarse map and its representation using a depth graph mutual information model (DGI), and also learns the mutual information between the graph to be predicted and its representation; it uses graph contamination level to determine whether the difference between the two mutual information values ​​is less than a threshold. If not, it selects multiple edges from the remaining perturbation edges and adds them to the graph to be predicted to generate a new graph to be predicted, and again determines whether the difference between the mutual information of the new graph to be predicted and its representation and the mutual information between the coarse map and its representation is less than a set threshold. If so, a clean graph is obtained.

[0053] The present invention also provides a computer device including a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to perform a graph data preprocessing method.

[0054] The present invention also provides a computer-readable storage medium storing a computer program adapted for loading by a processor to execute a graph data preprocessing method.

[0055] The above-described embodiments are merely preferred embodiments of the present invention, and the scope of protection of the present invention is not limited thereto. Any simple changes or equivalent substitutions of the technical solutions that can be obviously obtained by those skilled in the art within the scope of the technology disclosed in the present invention shall fall within the scope of protection of the present invention.

Claims

1. A graph data preprocessing method, characterized in that, Includes the following steps: Anomaly scores for nodes are obtained based on graph anomaly detection methods, and perturbation edges are determined based on the anomaly scores. Multiple perturbation edges are removed from the citation network dataset to be processed to obtain a rough graph; wherein, the citation network dataset is a graph data structure, nodes represent papers, and edges represent citation links between papers; Perturbation edges are used as candidate edges. Some edges are randomly selected from multiple candidate edges and added to the coarse map to generate the map to be predicted. The mutual information between the coarse map and the representation obtained from the coarse map is learned using the Depth Graph Mutual Information (DGI) model, as well as the mutual information between the map to be predicted and the representation obtained from the map to be predicted. The graph contamination level is used to determine whether the difference between the two mutual information values ​​is less than a threshold. If not, multiple edges are selected from the candidate edges and added to the map to be predicted to generate a new map to be predicted. The difference is then determined again to be less than a set threshold. If so, the coarse map is updated until a clean citation network is obtained.

2. The graph data preprocessing method according to claim 1, characterized in that, The process of removing multiple perturbation edges from the citation network dataset to obtain a rough map includes the following steps: Randomly select a node from the citation network dataset to be processed. i The Restart Random Walk (RWR) algorithm is used to process the nodes. i A random walk is performed, and the nodes and edges of the walk are combined to form positive samples. The collected positive samples are represented as positive sample subgraphs, which are groups of similar papers with the same theme and their citation links. Negative sample subgraphs are obtained by randomly selecting positive sample subgraphs from other nodes outside the positive samples. These negative sample subgraphs are groups of papers with completely different fields and their abnormal citation links. Scoring the positive and negative sample subplots separately includes the following steps: Learning nodes using the GCN model i The overall embedding representation of the subgraph; Using a deep neural network (DNN) to connect nodes i Mapping to the same embedding space as the overall embedding representation of the subgraph yields the nodes. i Embedded representation; Based on the overall embedding representation of the subgraph and nodes i The embedded representation is used, and the scores of the graph data nodes are obtained by combining the bilinear scoring function; The negative of the difference between the positive sample score and the negative sample score is taken as the outlier score; based on the graph data nodes. i The abnormal score derivation graph data is used to derive the abnormal scores of the edges, sort the abnormal scores of the edges, and remove the top ones. k By analyzing the edges, we can obtain a rough map.

3. The graph data preprocessing method according to claim 2, characterized in that, The abnormal scores of graph edges are derived from the abnormal scores of graph data nodes using the following formula: ； In the formula, S i For nodes i Abnormal scores, S j For nodes j Abnormal scores, A i，j =1 indicates a node i and nodes j There are edges between them. A i，j =0 indicates a node i and nodes j There is no edge between them.

4. The graph data preprocessing method according to claim 3, characterized in that, The learning nodes using the GCN model i The subgraph embedding representation is performed using the following formula: ； In the formula, k The subgraph represents the first k 1 node K Indicates the number of nodes in the subgraph. H k Indicates the first k The representation of each node.

5. The graph data preprocessing method according to claim 2, characterized in that, The node i The embedded representation is obtained by the following formula: ； In the formula, h i (l-1) Represents a node i In the l- The representation of layer 1, express l Parameters of layer -1 l Indicates the number of floors. It is the ReLU activation function. h i Table Node i The overall embedding representation of the subgraph.

6. The graph data preprocessing method according to claim 1, characterized in that, After obtaining the rough map, the graph contamination level is used to determine the difference between the two mutual information values, where the graph contamination level is: ； In the formula, G This is a rough sketch. G 'This is the image to be predicted.' I ( G , f θ ( G ))for G and f θ ( G Mutual information between them f θ ( G () is a rough representation. I ( G ', f θ ( G '))for G 'and f θ ( G Mutual information between ') f θ ( G ') represents the graph to be predicted, GPD( θ) This represents the difference in mutual information between the rough map and the map to be predicted.

7. A graph data preprocessing system, characterized in that, include: The rough map acquisition module is used to remove multiple perturbation edges from the graph data to be processed to obtain a rough map; The graph purification module uses perturbation edges as candidate edges, randomly selects some edges from multiple candidate edges and adds them to the rough map to generate the graph to be predicted; it uses the Depth Graph Mutual Information (DGI) model to learn the mutual information between the rough map and the representation obtained from the rough map, as well as the mutual information between the graph to be predicted and the representation obtained from the graph to be predicted. The graph contamination level is used to determine whether the difference between the two mutual information values ​​is less than a threshold. If not, multiple edges are selected from the candidate edges and added to the graph to generate a new graph to be predicted. The new difference is then determined again to be less than the set threshold. If so, the rough graph is updated until a clean citation network is obtained.

8. A computer device, characterized in that, It includes a memory and a processor; the memory stores a computer program, and the processor is used to run the computer program in the memory to perform the graph data preprocessing method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program adapted for loading by a processor to perform the graph data preprocessing method according to any one of claims 1-7.

Images