Abnormal user classification method based on graph neural network

CN119046775APending Publication Date: 2024-11-29NAT COMP NETWORK & INFORMATION SECURITY MANAGEMENT CENT
View PDF 0 Cites 1 Cited by

Patent Information

Application Number
CN202411131128.1
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-08-17
Publication Date
2024-11-29

AI Technical Summary

Benefits of technology

[0039] (1) The present invention uses the same matching information and mismatching information of nodes. This method can more accurately capture the role of nodes in the network, thereby improving the classification accuracy of the model, especially when identifying users with similar behavior patterns. This method can significantly reduce the false alarm rate; the signal aggregation method based on node-level homogeneity can accurately capture the role of nodes in their network structure. By learning to distinguish low-frequency signals from high-frequency signals, the model can effectively capture the homogamy and heterogamy characteristics in the network, thereby classifying nodes more accurately.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119046775A_ABST
    Figure CN119046775A_ABST
Patent Text Reader

Abstract

The invention relates to the technical field of information, and particularly discloses an abnormal user classification method based on a graph neural network, and the method comprises the steps: S1, data preprocessing and graph construction, S2, node representation decomposition, S3, generation of a pseudo label based on a matching degree matrix, S4, signal aggregation, S5, model design, S6, loss function design, and S7, a node classification method. According to the method, a Weisfeier-Lehman isomorphic test and syntactic degree matrix pseudo tag generation method is used, so that the training time and computing resources can be effectively reduced, and an efficient signal aggregation method also enables the reasoning process to be faster; through a method of decomposing node representation and pseudo label generation, the model is more transparent when making a decision, and how the model makes a judgment by using homology and heterology information can be understood, so that the interpretability and the precision of the model are improved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the technical field of network security, and in particular relates to an abnormal user classification method based on a graph neural network. Background Art

[0002] With the rapid development of the Internet, the number of users on social media and online platforms has increased dramatically, but it has also brought about various bad information and behaviors. Therefore, how to quickly and accurately identify and classify these abnormal users has become an urgent problem that Internet platforms need to solve. The abnormal user classification model based on Graph Neura lNetwork (GNN) is a method that uses graph neural network algorithms to identify and classify potentially harmful users in social media, online platforms or other network environments. This model achieves accurate classification of abnormal users by representing users and their relationships as a graph structure and using the powerful ability of GNN to learn and represent these relationships.

[0003] In the study of node classification, how to improve the accuracy of node classification is a long-term goal of scholars. In order to improve the interpretability and classification accuracy of the model, the inventor proposed an abnormal user classification method based on graph neural network, which improves the model's ability to classify abnormal users by studying the same degree relationship between nodes and decoupling the low-frequency and high-frequency information of nodes. Summary of the invention

[0004] The purpose of the present invention is to provide an abnormal user classification method based on graph neural network to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A method for classifying abnormal users based on graph neural network includes the following steps:

[0007] S1. Data preprocessing and graph construction: collect user association information and convert the association information into a graph structure, where nodes represent users, edges represent the association relationship between users, and edge weights represent the degree of user association;

[0008] S2. Node representation decomposition: The Weisfield-Lehman isomorphism test is used to decompose the node representation, and the node representation is decomposed into the similarity between the same information and the neighbors and the difference between the different information and the neighbors. Through the decomposed node representation, the node feature vector is extracted, which includes the node's own features and the neighbor relationship features;

[0009] S3, generate pseudo labels based on the same degree matrix, construct the same degree matrix, measure the similarity and difference between the node and its neighbors, and generate pseudo labels of the nodes according to the same degree matrix. The pseudo labels are used to divide the neighbor nodes into positive neighbors and negative neighbors;

[0010] S4, signal aggregation, uses the signal aggregation method to accurately capture the role of nodes in the network structure;

[0011] S5. Model design: Design a model based on graph neural network and use the final feature representation of nodes for training, including node update, data training and model optimization. Use pseudo labels and feature representation of nodes as training data and optimize model performance through cross-validation and hyperparameter tuning.

[0012] S6, loss function design, training by reducing the cross entropy error between the predicted and true labels;

[0013] S7. Abnormal user classification: Use the trained graph neural network model to classify the nodes in the graph, identify potential abnormal behaviors, and deploy the trained model to the actual system to perform real-time abnormal user classification.

[0014] Preferably, the node representation in step S2 is decomposed into:

[0015]

[0016]

[0017] Preferably, the iso-compatibility matrix in step S3 is expressed as:

[0018] P = Softmax(MLP(X))

[0019] C=PP T .

[0020] Preferably, in step S3, positive neighbors represent nodes that are similar to the central node, and negative neighbors represent nodes that are different from the central node. According to pseudo labels, the neighbor nodes are classified to study the influence of different categories of neighbor nodes on the central node.

[0021] Preferably, the aggregation process of the signal aggregation method in step S4 is:

[0022] Learning low-frequency signals: Capturing the assortativity in the network and learning similar features from positive neighbors;

[0023] Learning high-frequency signals: Capturing heterogeneity in the network and learning different features from negative neighbors;

[0024] Learning center node characteristics: no topological information is learned from the node itself.

[0025] Preferably, the signal aggregation method in step S4 is expressed as:

[0026] Learning low-frequency signals to capture the network's assimilation:

[0027]

[0028] Learning high-frequency signals to capture the heterogeneity of the network:

[0029]

[0030] Captures the structure-independent central node features:

[0031]

[0032] Therefore, the update equation of the node is:

[0033]

[0034] Preferably, the graph neural network based on step S5 includes GCN, GraphSAGE, and GAT.

[0035] Preferably, the loss function in step S6 is:

[0036]

[0037] Where yi is the true label, is the predicted label.

[0038] Compared with the prior art, the present invention has the following beneficial effects:

[0039] (1) The present invention utilizes the node's homogeneity information and heterogeneity information. This method can more accurately capture the role of the node in the network, thereby improving the classification accuracy of the model. In particular, when identifying users with similar behavior patterns, this method can significantly reduce the false alarm rate. The signal aggregation method based on node-level homogeneity can accurately capture the role of the node in its network structure. By distinguishing between low-frequency and high-frequency signal learning, the model can effectively capture the homogeneity and heterogeneity characteristics in the network, thereby more accurately classifying nodes.

[0040] (2) The present invention uses the Weisfielder-Lehman isomorphism test and the pseudo-label generation method of the homology matrix, which can effectively reduce the training time and computing resources; the efficient signal aggregation method also makes the reasoning process faster and is suitable for real-time detection scenarios; the decomposition node representation and pseudo-label generation method make the model more transparent when making decisions; the user can understand how the model uses homology and heterology information to make judgments, thereby increasing the interpretability of the model and the user's trust, and accurately identifying abnormal users. BRIEF DESCRIPTION OF THE DRAWINGS

[0041] Figure 1 This is a flow chart of the abnormal user classification method based on graph neural network of the present invention;

[0042] Figure 2 The present invention provides a block diagram of the social network user representation learning method based on representation decoupling. DETAILED DESCRIPTION

[0043] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments 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 creative work are within the scope of protection of the present invention.

[0044] Embodiment 1:

[0045] See also Figure 1 to Figure 2 As shown in FIG. 1 , the abnormal user classification method based on graph neural network includes the following steps:

[0046] S1. Data preprocessing and graph construction: collect user association data and convert the association data into a graph structure, where nodes represent users, edges represent the association relationship between users, and edge weights represent the degree of user association;

[0047] S2. Node representation decomposition: The Weisfield-Lehman isomorphism test is used to decompose the node representation, and the node representation is decomposed into the similarity between the same information and the neighbors and the difference between the different information and the neighbors. Through the decomposed node representation, the node feature vector is extracted, which includes the node's own features and the neighbor relationship features;

[0048] S3, generate pseudo labels based on the same degree matrix, construct the same degree matrix, measure the similarity and difference between the node and its neighbors, and generate pseudo labels of the nodes according to the same degree matrix. The pseudo labels are used to divide the neighbor nodes into positive neighbors and negative neighbors;

[0049] S4, signal aggregation, uses the signal aggregation method to accurately capture the role of nodes in the network structure;

[0050] S5. Model design: Design a model based on graph neural network and use the final feature representation of nodes for training, including node update, data training and model optimization. Use pseudo labels and feature representation of nodes as training data and optimize model performance through cross-validation and hyperparameter tuning.

[0051] S6, loss function design, training by reducing the cross entropy error between the predicted and true labels;

[0052] S7. Abnormal user classification: Use the trained graph neural network model to classify the nodes in the graph, identify potential abnormal behaviors, and deploy the trained model to the actual system to perform real-time abnormal user classification.

[0053] Specifically, the node representation in step S2 is decomposed into:

[0054]

[0055]

[0056] Specifically, the same matching matrix in step S3 is expressed as:

[0057] P = Softmax(MLP(X))

[0058] C=PP T .

[0059] Specifically, in step S3, positive neighbors represent nodes that are similar to the central node, and negative neighbors represent nodes that are different from the central node. According to pseudo labels, the neighbor nodes are classified to study the influence of different categories of neighbor nodes on the central node.

[0060] Specifically, the aggregation process of the signal aggregation method in step S4 is:

[0061] Learning low-frequency signals: Capturing the assortativity in the network and learning similar features from positive neighbors;

[0062] Learning high-frequency signals: Capturing heterogeneity in the network and learning different features from negative neighbors;

[0063] Learning center node characteristics: no topological information is learned from the node itself.

[0064] Specifically, the signal aggregation method in step S4 is expressed as:

[0065] Learning low-frequency signals to capture the network's assimilation:

[0066]

[0067] Learning high-frequency signals to capture the heterogeneity of the network:

[0068]

[0069] Captures the structure-independent central node features:

[0070]

[0071] Therefore, the update equation of the node is:

[0072]

[0073] Specifically, the graph neural network in step S5 includes GCN, GraphSAGE, and GAT.

[0074] Specifically, the loss function in step S6 is:

[0075]

[0076] Where yi is the true label, is the predicted label.

[0077] like Figure 2 As shown, nodes u1, u2, u3, u4 and u5 are positive neighbors, and node c1 learns low-frequency signals from nodes u1, u2, u3, u5 and u6; nodes c2, u7, u5 and u6 are negative neighbors, and c1 learns high-frequency signals from nodes c2, u7, u5 and u6; u5 and u6 are fuzzy nodes, participating in the positive and negative neighbor propagation of c1; c1 learns structure-independent features from itself.

[0078] As can be seen from the above, by using the node's homogeneity information and heterogeneity information, this method can more accurately capture the role of the node in the network, thereby improving the classification accuracy of the model, especially when identifying users with similar behavior patterns, this method can significantly reduce the false alarm rate; the signal aggregation method based on node-level homogeneity can accurately capture the role of the node in its network structure. By distinguishing between low-frequency and high-frequency signal learning, the model can effectively capture the homogeneity and heterogeneity characteristics in the network, thereby more accurately classifying nodes.

[0079] The use of the Weisfeiler-Lehman isomorphism test and pseudo-label generation method for homogeneity matrices can effectively reduce training time and computing resources. The efficient signal aggregation method also makes the reasoning process faster and suitable for real-time detection scenarios; the decomposition of node representation and pseudo-label generation methods make the model more transparent when making decisions. Users can understand how the model uses homogeneity and heterogeneity information to make judgments, thereby increasing the interpretability and accuracy of the model, and improving user experience and satisfaction by accurately identifying abnormal behaviors.

[0080] In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "examples", "specific examples" or "some examples" etc. means 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 may be combined in any one or more embodiments or examples in a suitable manner. In addition, those skilled in the art may combine and combine different embodiments or examples described in this specification and the features of different embodiments or examples, unless they are contradictory.

[0081] In the drawings of the embodiments disclosed in the present invention, only the structures related to the embodiments disclosed in the present invention are involved, and other structures can refer to the general design. In the absence of conflict, the same embodiment and different embodiments of the present invention can be combined with each other.

[0082] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A method for classifying abnormal users based on graph neural network, characterized in that: The following steps are involved: S1. Data preprocessing and graph construction: collect node data, including the relationship between nodes, and convert the relationship data into a graph structure, where nodes represent users, edges represent user relationships, and edge weights represent the degree of association; S2. Node representation decomposition: The Weisfeiler-Lehman isomorphism test is used to decompose the node representation, and the node representation is decomposed into the similarity between the same information and the neighbors and the difference between the different information and the neighbors. Through the decomposed node representation, the node feature vector is extracted, which includes the node's own features and the neighbor relationship features; S3, generate pseudo labels based on the same degree matrix, construct the same degree matrix, measure the similarity and difference between the node and its neighbors, and generate pseudo labels of the nodes according to the same degree matrix. The pseudo labels are used to divide the neighbor nodes into positive neighbors and negative neighbors; S4, signal aggregation, uses the signal aggregation method to accurately capture the role of nodes in the network structure; S5. Model design: Design a model based on graph neural network and use the final feature representation of nodes for training, including node update, data training and model optimization. Use pseudo labels and feature representation of nodes as training data and optimize model performance through cross-validation and hyperparameter tuning. S6, loss function design, training by reducing the cross entropy error between the predicted and true labels; S7. Abnormal user classification: Use the trained graph neural network model to classify the nodes in the graph, identify potential abnormal behaviors, and deploy the trained model to the actual system to perform real-time abnormal user classification detection.

2. According to claim 1, the abnormal user classification method based on graph neural network is characterized in that: The node representation in step S2 is decomposed into:

3. The abnormal user classification method based on graph neural network according to claim 1 is characterized in that: The same degree matrix in step S3 is expressed as: P = Softmax(MLP(X)) C=PP T 。 4. According to claim 1, the abnormal user classification method based on graph neural network is characterized in that: In step S3, positive neighbors indicate nodes that are similar to the central node, and negative neighbors indicate nodes that are different from the central node. Based on pseudo labels, the neighbor nodes are classified to study the impact of different categories of neighbor nodes on the central node.

5. According to claim 1, the abnormal user classification method based on graph neural network is characterized in that: The aggregation process of the signal aggregation method in step S4 is: Learning low-frequency signals: Capturing the assortativity in the network and learning similar features from positive neighbors; Learning high-frequency signals: Capturing heterogeneity in the network and learning different features from negative neighbors; Learning center node characteristics: no topological information is learned from the node itself.

6. According to claim 1, the abnormal user classification method based on graph neural network is characterized in that: The signal aggregation method in step S4 is expressed as: Learning low-frequency signals to capture the network's assimilation: Learning high-frequency signals to capture the heterogeneity of the network: Captures the structure-independent central node features: Therefore, the update equation of the node is:

7. The abnormal user classification method based on graph neural network according to claim 1 is characterized by: The graph neural networks in step S5 include GCN, GraphSAGE, and GAT.

8. The abnormal user classification method based on graph neural network according to claim 1 is characterized by: The loss function in step S6 is: Where yi is the true label, is the predicted label.

Citation Information

Cited By

  • Infrared small target detection method for graph structure analysis

    CN120451512A