A training method for a fake account detection model based on intra-class consistency hybrid

Abstract

This invention relates to a training method for a fake account detection model based on intra-class consistency mixing. The method first identifies strong and weak connections within the same class of nodes by calculating the feature similarity between account nodes, and constructs a perturbation mask matrix based on the weak connections. Simultaneously, node features are mixed according to similarity matching to generate graph data with multiple structural perturbations. Then, multi-order parameterless feature propagation is performed on each perturbation graph, and the results are aggregated. The obtained node representations are input into a multilayer perceptron to obtain the predicted probability distribution. Finally, by calculating the mean distribution of the predicted probabilities of all perturbation graphs and constraining the consistency between the predictions of each perturbation graph and the mean distribution, a joint optimization objective is constructed to train the model. This results in a fake account detection model that enhances intra-class representation diversity, effectively resists structural perturbations, maintains prediction stability, and improves detection accuracy.

Description

Technical Field

[0001] This invention belongs to the field of graph neural network and machine learning technology, and specifically relates to a training method for a fake account detection model based on intra-class consistency hybridization. Background Technology

[0002] With the rapid development of social networks, the problem of fake accounts has become increasingly serious. Fake accounts severely damage the ecosystem of social networks and pose a substantial threat to user information security, cyberspace governance, and platform reputation. Therefore, efficiently and accurately detecting fake accounts has become a critical task in the operation and security protection of social network platforms.

[0003] In existing technologies, graph structure data can naturally represent entities and their interactions in social networks, thus possessing significant application value in fake account detection. Graph neural networks, as the core method of graph representation learning, learn node representations by iteratively aggregating information from neighboring nodes. They can effectively capture the structural features and attribute information of graphs, achieving remarkable results in tasks such as node classification and link prediction, and have become one of the mainstream technical approaches for fake account detection in social networks.

[0004] However, existing graph neural networks still face several challenges. First, as the number of propagation layers increases, node representations gradually converge, leading to oversmoothing and a decrease in feature discrimination, thus weakening the model's discriminative ability. Second, graph-structured data often contains noise and perturbations, such as mutual following between fake accounts or maliciously added follower relationships. Traditional GNNs lack robustness to structural perturbations, resulting in unstable predicted distributions. Third, existing methods may disrupt key topological structures or introduce class conflicts during cross-class mixing or random dropout operations, limiting the model's generalization ability.

[0005] To address the aforementioned issues, existing technologies primarily employ residual connections, attention mechanisms, and jumper networks to mitigate oversmoothing, or enhance robustness through graph data augmentation. However, these methods typically lack theoretical guarantees regarding the controllability of perturbations and struggle to maintain intra-class diversity while avoiding cross-class mixing conflicts, resulting in lower detection accuracy and a higher likelihood of misclassification. Therefore, constructing a graph neural network method for detecting fake accounts on social networks that can enhance intra-class diversity, avoid cross-class conflicts, and simultaneously possess controllable perturbations and predictive consistency is a pressing technical problem that needs to be solved. Summary of the Invention

[0006] The purpose of this invention is to provide a training method for a fake account detection model based on intra-class consistency hybridization, which can enhance intra-class diversity, avoid cross-class conflicts, and have controllable perturbation, resulting in a detection model with high accuracy.

[0007] To achieve the above objectives, the technical solution adopted in this invention is a training method for a fake account detection model based on intra-class consistency hybrid model, which is carried out according to the following steps:

[0008] Step 1: Obtain training data, which includes... Individual account nodes and Each account node is used to build a graph neural network model based on the information obtained.

[0009] Step 2: Employ an intra-class consistency mixing mechanism to distinguish between intra-class strong and weak connections based on the feature similarity between account nodes. Construct a perturbation mask matrix based on the intra-class weak connections. Simultaneously, match nodes with strong correlations based on feature similarity to perform node mixing, generating multiple perturbation graphs. ;

[0010] Step 3: Perform multi-order parametric feature propagation on each perturbation graph, and aggregate the features of each propagation result to obtain the node representation;

[0011] Step 4: Input the aggregated node representation into the multilayer perceptron for probability prediction to obtain the predicted probability distribution of each perturbation map. Then, calculate the mean distribution of all predicted probabilities and calculate the KL divergence between each predicted probability and the mean of the predicted probabilities.

[0012] Step 5: Construct a joint optimization objective function that includes supervised loss and consistency loss. Update the model parameters through joint optimization. If the model parameters reach the convergence condition after the update, output the model. Otherwise, repeat steps 2 to 5 until the model reaches the convergence condition, and obtain the fake account detection model.

[0013] The construction of the graph neural network model in step 1 is carried out according to the following steps:

[0014] Step 1.1: Collect data containing... 1 node Account data of the strip edge data Establish a feature matrix for the node characteristics of account nodes. ,in Indicates the total number of nodes. This represents the feature dimensions of each node, including registration time, posting frequency, interaction behavior, and device information;

[0015] Then, corresponding label information is assigned according to the category of different nodes, and all the label information constitutes a label matrix. ,in This indicates the number of categories, including both legitimate accounts and fake accounts;

[0016] Establish an adjacency matrix based on the adjacency relationships between different nodes. Let A be a sparse matrix, when the nodes and nodes When there is an edge, i.e., when there is a relationship of interest, there exists. ,otherwise ;

[0017] Define adjacency matrix degree matrix The degree matrix is ​​a diagonal matrix whose diagonal elements satisfy... The degree matrix of the first degree The diagonal elements represent nodes. The degree, that is, the number of follow relationships that the user is connected to;

[0018] Step 1.2: For the adjacency matrix Normalization is performed to obtain the normalized adjacency matrix. :

[0019]

[0020] in, This is the identity matrix, used to add self-join terms;

[0021] Step 1.3: Based on the feature matrix Adjacency matrix Tag matrix Establish a graph neural network model.

[0022] The graph neural network model employs a multi-layer GNN based on message passing, with the inter-layer propagation rule as follows:

[0023]

[0024] in, Indicates the first Feature representation of layer nodes; It is the first The trainable weights of the layer; It is a non-linear activation function.

[0025] The generation of the perturbation map in step 2 is performed according to the following steps:

[0026] Step 2.1: For each tagged node Select other nodes of the same category to construct a set of nodes of the same type. Then calculate the points in that set. With each candidate node Feature similarity :

[0027] ;

[0028] Step 2.2: Based on the similarity score, assign each node... Select The most similar nodes of the same type are used to form an intra-class nearest neighbor set. :

[0029] ;

[0030] Step 2.3: For each node pair ,in Randomly sample a mixing coefficient This coefficient controls the fusion ratio of the two features to generate mixed sample features. and corresponding mixed tags :

[0031] , ;

[0032] Step 2.4: Set the calculation points With each candidate node Edge weight of the connecting edges Then calculate the average of all edge weights:

[0033] ;

[0034] Step 2.5: Use the average edge weights As a threshold for distinguishing between strong and weak connections, the edge weight is not less than the average edge weight. The connecting edge is a strong connecting edge. The edge weight is less than the average edge weight. The connecting edge is a weak connecting edge. ;

[0035] Step 2.6: Perturb the weak connections by randomly sampling a portion of edge data from the set of weak connections to construct the perturbed adjacency matrix. :

[0036]

[0037] in, , M is the perturbation matrix;

[0038] Step 2.7: Combine features from mixed samples Mixed tags and the perturbed adjacency matrix The perturbation diagram is obtained.

[0039] The node representation generation method in step 3 is performed according to the following steps:

[0040] Step 3.1: Perform parametric feature propagation on each perturbation map. The propagation formula is:

[0041]

[0042] in, For the first The first propagation after the first propagation A mixture of feature matrices, For the first The adjacency matrix of a perturbation graph For the first The first propagation after the first propagation The feature matrix of nodes, where In this method , , This indicates that the number of perturbation plots is set to between 3 and 8;

[0043] Step 3.2: Perform feature aggregation on the propagation results of each order to obtain the final node representation. :

[0044] .

[0045] Step 4 is the final representation of the nodes. The processing shall be carried out in accordance with the following steps:

[0046] Step 4.1: Finalize the node representation The data is input into a multilayer perceptron for probability prediction, yielding the predicted probability distribution for each perturbation map. :

[0047]

[0048] The Multilayer Perceptron (MLP) consists of an input layer, hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the node features, while the number of neurons in the hidden layer is typically set to 32 or 64. The output dimension is equal to the number of categories. Similarly, the output is converted into a probability distribution using the Softmax function;

[0049] Step 4.2: Calculate the mean distribution of the prediction probabilities for all perturbation maps:

[0050] ;

[0051] Step 4.3: Using the mean probability As pseudo-labels, calculate the predicted probability for each perturbation map. With mean probability KL divergence between:

[0052]

[0053] in, This indicates that the KL divergence is calculated node by node.

[0054] The construction of the fake account detection model in step 5 is carried out according to the following steps:

[0055] Step 5.1: Select the true label of a single node from the label matrix. ,Will The input is fed into a multilayer perceptron (MLP) to obtain the predicted label. Then define the supervision loss. Cross-entropy classification loss for labeled nodes:

[0056] ;

[0057] Step 5.2: Construct a joint optimization objective function, including supervised loss. With consistency loss :

[0058]

[0059] in, To control for consistency loss, the importance weighting coefficient is set between 0.5 and 2.0 in this method;

[0060] Step 5.3: Calculate the gradient of the model parameters using the backpropagation algorithm, minimize the joint optimization objective function using the optimizer, and update the model parameters;

[0061] Step 5.4: Input the validation set into the updated model for validation and calculate the validation loss. Then repeat steps 2 to 5 until the number of training iterations reaches 1000 or the rate of change of loss is less than 1e-4 in p consecutive training iterations. Terminate the training and output the model parameters at this time to obtain the fake account detection model.

[0062] Compared with the prior art, the present invention has the following advantages:

[0063] This invention proposes an intra-class consistency mixing mechanism to alleviate the class conflict problem caused by indiscriminate mixing in graph data in traditional mixing methods. It distinguishes between intra-class strong and weak connections based on node feature similarity, constructs a perturbation mask matrix based on the weak connections, and achieves feature mixing while preserving structure. Simultaneously, it matches nodes with strong correlations based on feature similarity for node mixing. This method enhances intra-class diversity while avoiding the contradictions caused by cross-class mixing.

[0064] This invention introduces a consistency constraint for multi-graph propagation, alleviating the problem of poor robustness of existing graph neural networks to structural noise. By performing parameter-free feature propagation under various structural perturbations, multi-view node embeddings are obtained, and KL divergence is used to constrain the consistency of prediction results under multiple perturbed graphs, enabling the model to learn a more generalizable feature space.

[0065] This invention separates feature propagation from feature transformation, alleviating the oversmoothing problem caused by deep propagation in traditional graph neural networks. It performs parameter-free feature propagation based on multi-structure perturbation, and then trains a classifier on the aggregated features, enabling the model to propagate higher-order features and reducing the risk of oversmoothing. Attached Figure Description

[0066] Figure 1 This is a flowchart of a training method for a fake account detection model based on intra-class consistency hybridization according to the present invention;

[0067] Figure 2 This is a schematic diagram of the intra-class consistency hybrid mechanism in this invention;

[0068] Figure 3 This is a schematic diagram illustrating the application of the present invention in detecting fake accounts on social networks. Detailed Implementation

[0069] The present invention will be further described below with reference to the accompanying drawings and specific embodiments.

[0070] Example 1:

[0071] like Figure 1 and Figure 2 As shown, the specific steps for training the fake account detection model using the method provided by this invention based on intra-class consistency hybridization are as follows:

[0072] Step 1: Collect data containing... 1 node Account data of the strip edge data Establish a feature matrix for the node characteristics of account nodes. ,in Indicates the total number of nodes. This represents the feature dimensions of each node, including registration time, posting frequency, interaction behavior, and device information;

[0073] Then, corresponding label information is assigned according to the category of different nodes, and all the label information constitutes a label matrix. ,in This indicates the number of categories, including both legitimate accounts and fake accounts;

[0074] Establish an adjacency matrix based on the adjacency relationships between different nodes. Let A be a sparse matrix, when the nodes and nodes When there is an edge, i.e., when there is a relationship of interest, there exists. ,otherwise ;

[0075] Define adjacency matrix degree matrix The degree matrix is ​​a diagonal matrix whose diagonal elements satisfy... The degree matrix of the first degree The diagonal elements represent nodes. The degree is the number of follow relationships that the user is connected to.

[0076] Step 2: Perform an adjacency analysis on the adjacency matrix. Normalization is performed to obtain the normalized adjacency matrix. :

[0077]

[0078] in, This is an identity matrix used to add self-join terms, ensuring that each node retains its own feature information during propagation and preventing its features from being diluted. The goal of normalization is to keep the feature values ​​of nodes stable during feature propagation, preventing numerical instability when high-degree nodes aggregate neighbor features.

[0079] Step 3: Based on the feature matrix Adjacency matrix Tag matrix Establish a graph neural network model.

[0080] Step 4: For each tagged node Select other nodes of the same category to construct a set of nodes of the same type. This ensures that subsequent node blending operations are performed only within the same category, thus avoiding the label semantic conflicts caused by cross-category blending in traditional blending methods. Then, points are calculated within this set. With each candidate node Feature similarity :

[0081]

[0082] A higher cosine similarity value indicates that the features of the two nodes are more similar. Cosine similarity is a measure of the similarity between two vectors in terms of direction. It is not affected by the vector magnitude, so it is particularly suitable for calculating the similarity of high-dimensional sparse feature vectors, such as user registration time, posting frequency, and interaction behavior in social networks.

[0083] Step 5: Based on the similarity score, assign each node... Select The most similar nodes of the same type are used to form an intra-class nearest neighbor set. :

[0084]

[0085] Using intra-class nearest neighbor set By constructing an intra-class nearest neighbor graph, each node is mixed only with nodes of the same class whose features are most similar to its own, thereby enhancing the diversity of intra-class data while maintaining the clarity of class boundaries.

[0086] Step 6: For each node pair ,in Randomly sample a mixing coefficient This coefficient controls the fusion ratio of the two features to generate mixed sample features. and corresponding mixed tags :

[0087] ,

[0088] By adjusting Controlling the distribution characteristics of mixing intensity;

[0089] Step 7: Set the calculation points With each candidate node Edge weight of the connecting edges Edge weights reflect the degree of feature similarity between nodes. Higher edge weights indicate greater similarity in features between the two nodes, and a more important connection between them; lower edge weights indicate greater feature difference between the two nodes, and a connection between them may contain noise or unreliable information. Then, the average of all edge weights is calculated:

[0090] .

[0091] Step 8: Use the average edge weight As a threshold for distinguishing between strong and weak connections, the edge weight is not less than the average edge weight. The connecting edge is a strong connecting edge. Strong connection edge This indicates connections between nodes that are highly similar in features, semantically related, and structurally reliable. The edge weight is less than the average edge weight. The connecting edge is a weak connecting edge. weak connection edge This indicates a connection where the nodes have significant differences in features, weak semantic relationships, and may be subject to noise risks.

[0092] Step 9: Perturb the weak connections by randomly sampling a portion of edge data from the weak connection set and constructing the perturbed adjacency matrix. :

[0093]

[0094] in, , M is the perturbation matrix. This perturbation strategy only affects weakly connected edges while fully preserving strongly connected edges, thereby enhancing data diversity while maintaining the key topological structure of the graph.

[0095] Step 10: Combine features from mixed samples Mixed tags and the perturbed adjacency matrix The perturbation diagram is obtained.

[0096] Step 11: Perform parameter-free feature propagation on each perturbation map. The propagation formula is:

[0097]

[0098] in, For the first The first propagation after the first propagation A mixture of feature matrices, For the first The adjacency matrix of a perturbation graph For the first The first propagation after the first propagation The feature matrix of nodes, where In this method , , The number of perturbation graphs is set to between 3 and 8. The propagation process does not introduce any trainable parameters, but only achieves the aggregation of neighbor features through the multiplication operation of the adjacency matrix and the feature matrix. Therefore, it is not affected by the training process and can stably propagate multi-order neighbor information.

[0099] Step 12: Perform feature aggregation on the propagation results of each order to obtain the final node representation. :

[0100]

[0101] This aggregation operation averages the results across all propagation orders. Through this average aggregation, the final representation of each node is a fusion of the results from itself to... Multi-scale feature information of neighbors. This aggregation strategy has the following advantages: First, it preserves feature information of different propagation orders, avoiding the oversmoothing problem that may be caused by using only the last propagation result; second, the average aggregation operation is simple and efficient, and does not require additional learnable parameters; finally, this aggregation method is equivalent to a multi-scale feature extraction mechanism, enabling the model to utilize both local and global structural information simultaneously.

[0102] Step 13: Finalize the node representation The data is input into a multilayer perceptron for probability prediction, yielding the predicted probability distribution for each perturbation map. :

[0103]

[0104] The Multilayer Perceptron (MLP) consists of an input layer, hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the node features, the number of neurons in the hidden layer is typically set to 64, and the output dimension is equal to the number of categories. Similarly, the output is converted into a probability distribution using the Softmax function.

[0105] Step 14: Calculate the mean distribution of the prediction probabilities for all perturbation maps:

[0106]

[0107] The mean distribution integrates the prediction information from all perturbation graphs and can be viewed as an integrated prediction of node categories. Since each perturbation graph only randomly perturbs weak connections while preserving the complete structure of strong connections, there are differences but also correlations between different perturbation graphs. By averaging the prediction results of multiple perturbation graphs, the prediction noise that a single perturbation graph might introduce can be reduced, resulting in more stable and reliable prediction results.

[0108] Step 15: Using the mean probability As pseudo-labels, calculate the predicted probability for each perturbation map. With mean probability KL divergence between:

[0109]

[0110] in, This indicates that the KL divergence is calculated node by node.

[0111] Step 16: Select the true label of a single node from the label matrix. ,Will The input is fed into a multilayer perceptron (MLP) to obtain the predicted label. Then define the supervision loss. Cross-entropy classification loss for labeled nodes:

[0112] ;

[0113] Step 17: Construct a joint optimization objective function, including supervised loss. With consistency loss :

[0114]

[0115] in, To control the importance weighting coefficient of consistency loss, it is set to between 0.5 in this method to balance the contributions of supervised learning and consistency regularization.

[0116] Step 18: Calculate the gradient of the model parameters using the backpropagation algorithm, minimize the joint optimization objective function using the optimizer, and update the model parameters;

[0117] Step 19: Input the validation set into the updated model for validation and calculate the validation loss. Then repeat steps 2 to 19 until the number of training iterations reaches 1000 or the rate of change of loss is less than 1e-4 in 100 consecutive training iterations. Terminate the training and output the model parameters at this time to obtain the fake account detection model.

[0118] To verify the effectiveness of the method of the present invention in detecting fake accounts on social networks, a system was constructed as follows: Figure 3 The diagram illustrates a simulated social network environment. This environment includes normal user nodes and fake account nodes, with follow-relationship edges between them. Strong connections represent trusted connections with mutual following and high-frequency interaction, and are preserved intact; weak connections represent unreliable connections with one-way following and low interaction, and are randomly perturbed to generate multiple views. In the diagram, blue square nodes represent labeled samples, and orange dashed arrows indicate successfully detected paths. Experimental results show that the method of this invention can effectively propagate information along the detection path from a small number of labeled samples, ultimately accurately identifying unlabeled fake accounts such as F4, verifying the effectiveness of the intra-class consistency hybrid mechanism and multi-graph propagation consistency constraints in detecting fake accounts on social networks.

[0119] In addition, four detection tasks were set up to test the model trained by this method. Each task had different requirements regarding the proportion of fake accounts, network density, and number of labeled samples. The specific task settings are as follows:

[0120] Task 1 5000 5% 20 0.008 Task 2 5000 10% 20 0.008 Task 3 10000 5% 20 0.006 Task 4 10000 15% 20 0.005

[0121] The detection model obtained in Example 1, along with the existing GCN and GAT detection models, were used for the detection of the above four detection tasks. The detection accuracy is shown in the table below:

[0122] Task 1 82.3% 84.1% 88.6% Task 2 78.5% 80.2% 86.2% Task 3 81.2% 83.0% 87.9% Task 4 75.8% 77.5% 84.7%

[0123] As shown in the table above, the detection model obtained by the training method provided in this application performs best on all tasks, which indicates that the introduction of intra-class consistency hybrid mechanism and multi-graph propagation consistency constraint effectively improves the performance of fake account detection.

Claims

1. A training method for a fake account detection model based on intra-class consistency hybrid model, characterized in that, Follow these steps: Step 1: Obtain training data, the training data including Individual account nodes and Each account node is used to build a graph neural network model based on the information obtained. Step 2: Employ an intra-class consistency mixing mechanism to distinguish between intra-class strong and weak connections based on the feature similarity between account nodes. Construct a perturbation mask matrix based on the intra-class weak connections. Simultaneously, match nodes with strong correlations based on feature similarity to perform node mixing, generating multiple perturbation graphs. ; Step 3: Perform multi-order parametric feature propagation on each perturbation graph, and aggregate the features of each propagation result to obtain the node representation; Step 4: Input the aggregated node representation into the multilayer perceptron for probability prediction to obtain the predicted probability distribution of each perturbation map. Then, calculate the mean distribution of all predicted probabilities and calculate the KL divergence between each predicted probability and the mean of the predicted probabilities. Step 5: Construct a joint optimization objective function that includes supervised loss and consistency loss. Update the model parameters through joint optimization. If the model parameters reach the convergence condition after the update, output the model. Otherwise, repeat steps 2 to 5 until the model reaches the convergence condition, and obtain the fake account detection model.

2. The method for training a fake account detection model based on intra-class consistency hybrid as described in claim 1, characterized in that, The construction of the graph neural network model in step 1 is carried out according to the following steps: Step 1.1: Collect data containing... 1 node Account data of the strip edge data Establish a feature matrix for the node characteristics of account nodes. ,in Indicates the total number of nodes. This represents the feature dimensions of each node, including registration time, posting frequency, interaction behavior, and device information; Then, corresponding label information is assigned according to the category of different nodes, and all the label information constitutes a label matrix. ,in This indicates the number of categories, including both legitimate accounts and fake accounts; Establish an adjacency matrix based on the adjacency relationships between different nodes. Let A be a sparse matrix, when the nodes and nodes When there is an edge, i.e., when there is a relationship of interest, there exists. ,otherwise ; Define adjacency matrix degree matrix The degree matrix is ​​a diagonal matrix whose diagonal elements satisfy... The degree matrix of the first degree The diagonal elements represent nodes. The degree, that is, the number of follow relationships that the user is connected to; Step 1.2: For the adjacency matrix Normalization is performed to obtain the normalized adjacency matrix. : in, This is the identity matrix, used to add self-join terms; Step 1.3: Based on the feature matrix Adjacency matrix Tag matrix Establish a graph neural network model.

3. The method for training a fake account detection model based on intra-class consistency hybrid as described in claim 2, characterized in that, The graph neural network model employs a multi-layer GNN based on message passing, with the inter-layer propagation rule as follows: in, Indicates the first Feature representation of layer nodes; It is the first The trainable weights of the layer; It is a non-linear activation function.

4. The method for training a fake account detection model based on intra-class consistency hybrid as described in claim 1, characterized in that, The generation of the perturbation map in step 2 is performed according to the following steps: Step 2.1: For each tagged node Select other nodes of the same category to construct a set of nodes of the same type. Then calculate the points in that set. With each candidate node Feature similarity : ； Step 2.2: Based on the similarity score, assign each node... Select The most similar nodes of the same type are used to form an intra-class nearest neighbor set. : ； Step 2.3: For each node pair ,in Randomly sample a mixing coefficient This coefficient controls the fusion ratio of the two features to generate mixed sample features. and corresponding mixed tags : ， ； Step 2.4: Set the calculation points With each candidate node Edge weight of the connecting edges Then calculate the average of all edge weights: ； Step 2.5: Use the average edge weights As a threshold for distinguishing between strong and weak connections, the edge weight is not less than the average edge weight. The connecting edge is a strong connecting edge. The edge weight is less than the average edge weight. The connecting edge is a weak connecting edge. ; Step 2.6: Perturb the weak connections by randomly sampling a portion of edge data from the set of weak connections to construct the perturbed adjacency matrix. : in, , M is the perturbation matrix; Step 2.7: Combine features from mixed samples Mixed tags and the perturbed adjacency matrix The perturbation diagram is obtained.

5. The method for training a fake account detection model based on intra-class consistency hybrid as described in claim 1, characterized in that, The node representation generation method in step 3 is performed according to the following steps: Step 3.1: Perform parametric feature propagation on each perturbation map. The propagation formula is: in, For the first The first propagation after the first propagation A mixture of feature matrices, For the first The adjacency matrix of a perturbation graph For the first The first propagation after the first propagation The feature matrix of nodes, where In this method , , This indicates that the number of perturbation plots is set to between 3 and 8; Step 3.2: Perform feature aggregation on the propagation results of each order to obtain the final node representation. : 。 6. The method for training a fake account detection model based on intra-class consistency hybrid as described in claim 1, characterized in that, In step 4, the nodes are finally represented. The processing shall be carried out in accordance with the following steps: Step 4.1: Finalize the node representation The data is input into a multilayer perceptron for probability prediction, yielding the predicted probability distribution for each perturbation map. : The Multilayer Perceptron (MLP) consists of an input layer, hidden layers, and an output layer. The number of neurons in the input layer is equal to the dimension of the node features, while the number of neurons in the hidden layer is typically set to 32 or 64. The output dimension is equal to the number of categories. Similarly, the output is converted into a probability distribution using the Softmax function; Step 4.2: Calculate the mean distribution of the prediction probabilities for all perturbation maps: ； Step 4.3: Using the mean probability As pseudo-labels, calculate the predicted probability for each perturbation map. With mean probability KL divergence between: in, This indicates that the KL divergence is calculated node by node.

7. The method for training a fake account detection model based on intra-class consistency hybrid as described in claim 1, characterized in that, The construction of the fake account detection model in step 5 is carried out according to the following steps: Step 5.1: Select the true label of a single node from the label matrix. ,Will The input is fed into a multilayer perceptron (MLP) to obtain the predicted label. Then define the supervision loss. Cross-entropy classification loss for labeled nodes: ； Step 5.2: Construct a joint optimization objective function, including supervised loss. With consistency loss : in, To control for consistency loss, the importance weighting coefficient is set between 0.5 and 2.0 in this method; Step 5.3: Calculate the gradient of the model parameters using the backpropagation algorithm, minimize the joint optimization objective function using the optimizer, and update the model parameters; Step 5.4: Input the validation set into the updated model for validation and calculate the validation loss. Then repeat steps 2 to 5 until the number of training iterations reaches 1000 or the rate of change of loss is less than 1e-4 in p consecutive training iterations. Terminate the training and output the model parameters at this time to obtain the fake account detection model.

Images