Social robot detection method based on KAN graph neural network model

Abstract

The invention provides a social robot detection method based on a KAN graph neural network model, and the method comprises the steps: carrying out the data preprocessing of an identified social account data set, carrying out the adaptive synthesis sampling through employing an ADASYN method, and carrying out the balance processing of the data set; after the balanced data set is input into the model, graph structure data is processed through two layers of convolution, the first layer extracts preliminary features, the second layer further learns multi-level graph features, in each layer of convolution, a KAN linear layer enhances the nonlinear expression ability of the features by using B spline interpolation, and the feature extraction efficiency is improved; the model is helped to better capture various social interaction type relationships in the graph data; neighborhood between nodes is used for learning representation of the nodes, the nodes with higher representation capability can be learned through two convolutional layers for training, classification prediction is further performed through a softmax function, and a final prediction result is obtained, so that social robot detection is realized.

Description

Technical Field

[0001] The present invention relates to the field of social robot detection, and in particular to a social robot detection method based on a KAN graph neural network model. Background Art

[0002] With the booming development of the Internet and information technology, social networks have attracted a large number of users and become an important part of today's cyberspace. Since the emergence of social networks, network social robots have been born with social networks. With the rise of social media platforms such as Facebook, Twitter, Instagram, etc., people socialize, share information and opinions on these platforms. However, these platforms have also become places for robot accounts to breed and operate. Nowadays, social networks have been deeply integrated into everyone's life. Platforms such as Facebook and Weibo already have hundreds of millions of active users. While they have amazing promotion and dissemination effects, they also provide opportunities for network social robots to maliciously manipulate information dissemination.

[0003] Social bots, also known as social accounts controlled by automated programs, have become a well-known phenomenon among social media users. Social bots pose a threat and challenge to online communities by spreading misinformation. Therefore, a lot of research work has been done to combat the malicious influence of social bots through automatic bot detection systems, case studies for specific situations, etc. In the field of social bot detection, current methods can be roughly divided into three categories: content-based methods, behavior-based methods, and social network structure-based methods. Content-based methods identify bots by analyzing the content of information posted by users, and these methods usually rely on natural language processing technology. Behavior-based methods detect bots by analyzing users' behavioral patterns, such as posting frequency, interaction frequency, etc. Social network structure-based methods use users' relationships and interaction patterns in social networks for identification.

[0004] A major challenge in social robot detection is data imbalance. In real social networks, the number of robot users is usually small compared to normal users, which leads to a serious imbalance in the ratio of positive and negative samples in the training data. When dealing with such imbalanced data, traditional machine learning algorithms tend to favor the majority class, resulting in poor detection results. Therefore, how to effectively deal with the problem of data imbalance and improve the accuracy and reliability of social robot detection. In addition, how to effectively capture the complex relationships and high-dimensional graph structure information in social networks is also a difficult problem for social robot detection. Users in social networks interact in a variety of ways, including simple attention relationships and more complex interaction patterns such as comments, forwarding, and citations. This multi-level interactive relationship is difficult to fully represent through traditional graph models; when dealing with these complex relationships, traditional methods can often only capture local or shallow graph features, making it difficult for the model to effectively distinguish between normal users and robot users; therefore, how to use graph neural networks to extract and enhance these complex graph relationships has become a major difficulty in social robot detection. Summary of the invention

[0005] The present invention provides a social robot detection method based on a KAN graph neural network model. The present invention processes unbalanced data sets through the ADASYN algorithm, and combines the advantages of the graph convolutional neural network with the KAN neural network to effectively improve the accuracy and robustness of social robot detection. See the following description for details:

[0006] A social robot detection method based on a KAN graph neural network model, characterized in that the method comprises: specific steps of:

[0007] Step 1: Use the ADASYN method to preprocess the minority samples in the unbalanced social robot dataset;

[0008] Step 2: These input data will be fed into the model, which combines the advantages of graph convolutional neural network and KAN neural network, and can process complex graph structure data and effectively capture nonlinear relationships;

[0009] Step 3: Each time the graph neural network performs message passing and aggregation, it first performs a nonlinear feature mapping through KAN, and then uses these KAN-transformed features for message passing and aggregation;

[0010] Step 4: In order to prevent the model from overfitting during training, the Dropout mechanism is used to randomly inactivate node features after each layer of the model is output;

[0011] Step 5. After feature processing by multi-layer graph convolution and KAN linear layer, each node will eventually generate a vector representing the features of the node. In order to perform classification tasks, the model uses the softmax function to calculate the probability distribution of the category to which each node belongs.

[0012] Furthermore, extracting data in step 1 specifically includes:

[0013] (1) Node features: including feature tensors extracted from descriptions, tweet-related features, numerical attribute features, which represent the numerical features of nodes, such as the number of followers and tweets of a user, and category attribute features, which represent the category features of nodes;

[0014] (2) Edge index: used to describe the connection relationship between nodes and indicate the interaction between nodes;

[0015] (3) Edge type: represents different types of edge relationships, including: follow type, friend type, mention type, reply type, reference type, link, and topic tag;

[0016] (4) Node label: 0 represents a normal user and 1 represents a social robot account.

[0017] Furthermore, D min represents the sample set of the minority class, D maj Represents the sample set of the majority class, including N min and N maj samples, for each minority class sample x i ∈D min , use the k nearest neighbor algorithm to find its nearest k neighbors, the neighbors are from the minority class, synthesize more training data for the minority class samples that are more difficult to learn, use the Euclidean distance metric, the number of synthetic samples generated for each minority class sample is based on the complexity of its neighbors, and define the total number of synthetic samples G that need to be generated as:

[0018] G=β*(N maj -N min )

[0019] Where β∈[0,1] is a defined balancing factor that controls the number of synthetic samples generated. Then, for each minority class sample x i The number of samples generated g i According to the formula:

[0020]

[0021] Among them, n i is the number of neighbors, for each minority class sample x i , according to the calculated g i, randomly select a neighbor x from its neighbors j , generate new synthetic samples by linear interpolation:

[0022] x new =x i +λ*(x j -x i )

[0023] Among them, λ∈[0,1] is a random number used to randomly generate new samples. This formula represents the generated synthetic sample x new Between x i and neighbor samples x j After that, the generated synthetic sample x new Added to the minority class sample set, we finally get a balanced social robot dataset.

[0024] Furthermore, the model in step 2 combines the advantages of graph convolutional neural networks and KAN neural networks, and can process complex graph structure data and effectively capture nonlinear relationships. One of its core components is the graph convolutional layer, which is used to aggregate the features of each node and its neighboring nodes and generate new node feature representations.

[0025] Further, the data pre-processed in step 1 is input into the invention model;

[0026] After two convolutional layers, the core calculation formula is:

[0027]

[0028] Among them, H (l) is the node feature representation of the lth layer, A is the original adjacency matrix, D is the degree matrix, is the adjacency matrix after adding self-loops, W (l) is the trainable weight matrix of the lth layer, σ is the nonlinear activation function. This step normalizes the adjacency matrix A, performs weighted aggregation of the features of the node and its neighbors in each layer, and updates the representation of the node.

[0029] Furthermore, in step 3, the input features are transformed nonlinearly through the KAN linear layer to generate new features after B-spline interpolation. The output of the KAN linear layer is then used as part of the node features. Next, these new features are fused with the adjacency matrix features for message passing and aggregation. In this way, each layer of GCN will combine the output of the KAN layer to update the representation of the node.

[0030] Furthermore, the KAN linear layer combines two parts: the basic part based on linear transformation and the nonlinear part based on B-spline interpolation. The KAN linear layer enhances the nonlinear mapping capability of features through B-spline interpolation. Its main function is to introduce a more flexible nonlinear transformation to replace the traditional linear mapping. B-spline interpolation is to calculate the B-spline basis function on a specific grid. The KAN linear layer realizes the local nonlinear interpolation of the input features, so that the features of each node can better adapt to complex relationships.

[0031] Its main formula is:

[0032]

[0033] Among them, f(x1,...,x n ) represents an n-dimensional continuous function, representing the input multidimensional data, Φ q represents a set of trainable one-dimensional functions for nonlinear transformation of the one-dimensional weighted sum of inputs, φ q,p (x p ) is another set of trainable one-dimensional functions that transform each input dimension x p transform; this formula shows that a multivariate function can be fundamentally reduced to a combination of appropriately defined univariate functions, where the combination involves only simple addition.

[0034] Furthermore, in step 4, in order to prevent the model from overfitting during the training process, the Dropout mechanism is used to randomly inactivate the node features after each layer of the model is output, and the Dropout mechanism is used to randomly inactivate the features after each layer of the model is output, so as to avoid overfitting of the model;

[0035] After the output of each layer, the Dropout mechanism is used to randomly inactivate the features to avoid overfitting of the model:

[0036] H (l+1) = Dropout(H (l+1) ,p)

[0037] Among them, p is the probability of Dropout, which means randomly discarding some features;

[0038] Furthermore, in step 5, in order to perform the classification task, the model uses the softmax function to calculate the probability distribution of the category to which each node belongs, and the nodes are finally classified by the softmax function.

[0039] Furthermore, after multi-layer graph convolution and feature processing of the KAN linear layer, each node will eventually generate a vector representing the node's features; in order to perform classification tasks, the model uses the softmax function to calculate the probability distribution of the category to which each node belongs;

[0040] Finally, classification is performed through the softmax function:

[0041] For the GCN classifier, the formula is as follows:

[0042]

[0043] in, is the output of the model, i.e., the probability distribution of the predicted label of each node; A is the adjacency matrix, which represents the connection relationship between nodes in the graph; H is the node feature matrix; softmax is the activation function, which converts the output into a probability distribution;

[0044] To simplify the calculation, the present invention uses the unit matrix E as the input adjacency matrix A; the label and enhancement embedding matrix Input into the formula for training;

[0045] The formula is as follows:

[0046]

[0047] It reduces the complexity of implementation and the difficulty of model design and implementation, making the model more efficient and easier to deploy. It also improves the training and prediction speed of the model, allowing the model to be applied to actual scenarios more quickly.

[0048] The final objective function of the classifier of the present invention can be written as:

[0049]

[0050] Among them, L(y') is the total loss function, y i is node v i The true label, y' i is the corresponding predicted label; V L is the set of nodes containing the true labels, V S is a set containing synthetic nodes; λ is a weight parameter that balances the loss of real nodes and synthetic nodes, ranging from 0 to 1; the loss function measures the supervision loss between the real label and the predicted label.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] The present invention proposes a social robot detection method based on a graph neural network model of KAN, which performs data preprocessing on the identified social account data set, uses the ADASYN method for adaptive synthetic sampling, and balances the data set; after the balanced data set is input into the model, the graph structure data is processed through two layers of convolution, the first layer extracts preliminary features, and the second layer further learns multi-level graph features. In each layer of convolution, the KAN linear layer uses B-spline interpolation to enhance the nonlinear expression ability of the features, helping the model to better capture various types of social interaction relationships in the graph data; the neighbor relationship between nodes is used to learn the representation of the nodes, and through two layers of convolution layers, nodes with more representation capabilities can be learned for training, and further classification prediction is performed through the softmax function to obtain the final prediction result, thereby realizing social robot detection.

[0053] The ADASYN preprocessing method adopted in the present invention enhances the model's recognition ability for samples that are difficult to classify. When the social robot faces diverse user inputs, it can better solve the data bias problem. At the same time, unbalanced data may cause the social robot model to be biased, making it overly dependent on the majority class and ignoring the performance of the minority class. The ADASYN preprocessing method adopted can improve the classifier's recognition ability for the minority class.

[0054] The present invention adopts a new neural network architecture, implements learnable spline-based functions between layers, changes the way of processing information across graph structures, effectively improves the model's ability to capture data patterns in the data set, and decomposes complex multivariate problems into multiple simple single-variable functions; this decoupling method can analyze the impact of each variable on the output separately, and thus more clearly explain the behavior of the model, while providing powerful nonlinear mapping capabilities, and can perform fine-grained nonlinear processing of features through learnable single-variable functions, thereby enhancing the model's ability to capture multi-level features and increasing the model's flexibility, enabling it to better fit a variety of social interaction relationships between different user groups, and improving the model's generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0055] Figure 1 This is a flowchart of a social robot detection method based on the KAN graph neural network model;

[0056] Figure 2 This is a flow chart of a model training method for a social robot detection method based on a KAN graph neural network model. DETAILED DESCRIPTION

[0057] In order to make the objectives, technical solutions and advantages of the present invention more clear, the embodiments of the present invention are described in further detail below.

[0058] Example 1

[0059] The present invention provides a social robot detection method based on a graph neural network model of KAN, performs data preprocessing on the identified social account data set, uses the ADASYN method to adaptively synthesize sampling, and balances the data set; after the balanced data set is input into the model, the graph structure data is processed through two layers of convolution, the first layer extracts preliminary features, and the second layer further learns multi-level graph features. In each layer of convolution, the KAN linear layer uses B-spline interpolation to enhance the nonlinear expression ability of the features, helping the model to better capture various types of social interaction relationships in the graph data; the neighbor relationship between nodes is used to learn the representation of the nodes, and nodes with more representation capabilities can be learned through two layers of convolution layers for training, and further classification prediction is performed through the softmax function to obtain the final prediction result to achieve social robot detection; the method includes the following steps:

[0060] Step 1: For the unbalanced social robot dataset, use the ADASYN method to preprocess the data; the dataset before preprocessing includes:

[0061] (1) Node features: including feature tensors extracted from descriptions, tweet-related features, numerical attribute features, which represent the numerical features of nodes, such as the number of users’ followers, the number of tweets, and other numerical data, and category attribute features, which represent the category features of nodes.

[0062] (2) Edge index: used to describe the connection relationship between nodes and indicate the interaction between nodes.

[0063] (3) Edge type: represents different types of edge relationships, which is very important in social networks because edges may represent different interaction types or hierarchical relationships.

[0064] (4) Node label: 0 represents a normal user and 1 represents a social robot account.

[0065] Data imbalance refers to the large difference between the ratio of minority class samples and majority class samples, which will cause the machine learning model to favor majority class samples, thus affecting the generalization ability of the model. The core idea of ​​this method is to adaptively synthesize new samples for minority class samples according to the learning difficulty of minority class samples. ADASYN focuses on those minority class samples that are difficult to learn and generates more synthetic samples to improve their learning effect.

[0066] Among them, D min represents the sample set of the minority class, D maj Represents the sample set of the majority class, including N min and N maj samples. For each minority class sample x i∈D min , use the k nearest neighbor algorithm to find its nearest k neighbors, the neighbors are from the minority class, synthesize more training data for the minority class samples that are more difficult to learn, use the Euclidean distance metric, the number of synthetic samples generated for each minority class sample is based on the complexity of its neighbors, and define the total number of synthetic samples G that need to be generated as:

[0067] G=β*(N maj -N min )

[0068] Among them, β∈[0,1] is a defined balancing factor that controls the number of synthetic samples generated. Then, for each minority class sample x i The number of samples generated g i According to the formula:

[0069]

[0070] Among them, n i is the number of neighbors, for each minority class sample x i , according to the calculated g i , randomly select a neighbor x from its neighbors j , generate new synthetic samples by linear interpolation:

[0071] x new =x i +λ*(x j -x i )

[0072] Among them, λ∈[0,1] is a random number used to randomly generate new samples; this formula represents the generated synthetic sample x new Between x i and neighbor samples x j After that, the generated synthetic sample x new Add to the minority class sample set, and finally get a balanced social robot dataset;

[0073] Step 2: These input data will be fed into the model, which combines the advantages of graph convolutional neural network and KAN neural network, and can process complex graph structure data and effectively capture nonlinear relationships;

[0074] One of the core components of the invented model is the graph convolution layer, which is used to aggregate the features of each node with its neighboring nodes and generate a new node feature representation; this process is calculated in each layer using the following formula:

[0075]

[0076] Among them, H (l)is the node feature representation of the lth layer, A is the original adjacency matrix, D is the degree matrix, is the adjacency matrix after adding self-loops, W (l) is the trainable weight matrix of the lth layer, σ is the nonlinear activation function. This step normalizes the adjacency matrix A, performs weighted aggregation of the features of the node and its neighbors in each layer, and updates the representation of the node.

[0077] Step 3: Each time the graph neural network performs message passing and aggregation, it first performs a nonlinear feature mapping through KAN, and then uses these KAN-transformed features for message passing and aggregation;

[0078] In the forward propagation process of GCN, the features are linearly transformed through the linear layer, and then further processed through the KAN linear layer. The input features are further nonlinearly transformed through KAN to generate new features after B-spline interpolation; the output of the KAN linear layer is used as part of the node features. Next, these new features are multiplied with the adjacency matrix for message passing and aggregation; in this way, each layer of GCN combines the output of the KAN layer to update the representation of the node.

[0079] The KAN linear layer enhances the nonlinear mapping capability of features through B-spline interpolation. Its main function is to introduce a more flexible nonlinear transformation to replace the traditional linear mapping. B-spline interpolation is to calculate the B-spline basis function on a specific grid. The KAN linear layer realizes the local nonlinear interpolation of the input features, so that the features of each node can better adapt to complex relationships. Its main formula is:

[0080]

[0081] Among them, f(x1,...,x n ) represents an n-dimensional continuous function, representing the input multidimensional data, Φ q represents a set of trainable one-dimensional functions for nonlinear transformation of the one-dimensional weighted sum of inputs, φ q,p (x p ) is another set of trainable one-dimensional functions that transform each input dimension x p Transformation; This formula shows that a multivariate function can be fundamentally reduced to a combination of properly defined univariate functions, where the combination involves only simple addition; with the help of KAN's nonlinear mapping to handle more complex feature relationships, the performance of the model on complex tasks is improved.

[0082] Step 4: In order to prevent the model from overfitting during training, the Dropout mechanism is used to randomly inactivate node features after each layer of the model is output;

[0083] After the output of each layer, the Dropout mechanism is used to randomly inactivate the features to avoid overfitting of the model:

[0084] H (l+1) = Dropout(H (l+1) ,p)

[0085] Among them, p is the probability of Dropout, which means randomly discarding some features. In each forward propagation, Dropout randomly discards some neurons to make the model more robust; this can avoid the model's over-reliance on certain specific features, thereby improving the model's generalization ability.

[0086] Step 5: After multi-layer graph convolution and feature processing of KAN linear layer, each node will eventually generate a vector representing the node feature;

[0087] In order to perform classification tasks, the model uses the softmax function to calculate the probability distribution of the category to which each node belongs. Finally, classification is performed through the softmax function:

[0088] For the GCN classifier, the formula is as follows:

[0089]

[0090] in, is the output of the model, that is, the probability distribution of the predicted label of each node; A is the adjacency matrix, which represents the connection relationship between nodes in the graph; H is the node feature matrix; softmax is the activation function, which converts the output into a probability distribution.

[0091] To simplify the calculation, the present invention uses the identity matrix E as the input adjacency matrix A. The labels and enhancements are embedded in the matrix Input into the formula for training;

[0092] The formula is as follows:

[0093]

[0094] It reduces the complexity of implementation and the difficulty of model design and implementation, making the model more efficient and easier to deploy. It also improves the training and prediction speed of the model, allowing the model to be applied to actual scenarios more quickly.

[0095] The final objective function of the classifier of the present invention can be written as:

[0096]

[0097] Among them, L(y') is the total loss function, y i is node vi The true label, y' i is the corresponding predicted label; V L is the set of nodes containing the true labels, V S is a set containing synthetic nodes; λ is a weight parameter that balances the loss of real nodes and synthetic nodes, ranging from 0 to 1; the loss function measures the supervision loss between the real label and the predicted label.

[0098] The introduction of the weight parameter λ provides flexibility so that the loss function can be adjusted according to specific circumstances. By adjusting the value of λ, the relative importance of the real node loss and the synthetic node loss can be flexibly controlled to optimize the model performance.

[0099] In summary, Example 1 of the present invention realizes the prediction of social robots using graph convolutional neural networks. By utilizing ADASYN to process unbalanced data sets and combining the advantages of graph convolutional neural networks and KAN neural networks, the accuracy and robustness of social robot detection are effectively improved. The present invention not only improves the accuracy and efficiency of the social robot detection algorithm, but also enhances its generalization ability on different data sets, thereby realizing efficient social robot detection.

[0100] Experimental Example 2

[0101] Based on the social robot detection method based on the KAN graph neural network model provided above, this test case uses the MGTAB dataset, which contains 10,384 users annotated by experts, including 7,554 human users and 2,830 social robots. The experimental results are as follows:

[0102]

[0103] From the data, it can be concluded that the social robot detection method based on the KAN graph neural network model provided by the invention can predict robots to a certain extent and improve the accuracy; the data shows that the present invention can give correct classification results in more samples, and the model performs better in overall performance; the improvement in F1 score shows that the present invention can not only accurately predict positive samples, but also effectively process negative samples, and the performance is more balanced; at the same time, the improvement in recall rate means that the present invention can identify more positive samples, which is very important for improving the accuracy and reliability of social robot detection, provides important guarantees for the healthy maintenance of social platforms, prevents robot accounts from interfering with platform content, and helps to establish a more real and credible social environment.

Claims

1. A social robot detection method based on a KAN graph neural network model, characterized in that: The method comprises the following specific steps: Step 1: Use the ADASYN method to preprocess the minority samples in the unbalanced social robot dataset; Step 2: These input data will be fed into the model, which combines the advantages of graph convolutional neural network and KAN neural network, and can process complex graph structure data and effectively capture nonlinear relationships; Step 3: Each time the graph neural network performs message passing and aggregation, it first performs a nonlinear feature mapping through KAN, and then uses these KAN-transformed features for message passing and aggregation; Step 4: In order to prevent the model from overfitting during training, the Dropout mechanism is used to randomly inactivate node features after each layer of the model is output; Step 5. After feature processing by multi-layer graph convolution and KAN linear layer, each node will eventually generate a vector representing the features of the node. In order to perform classification tasks, the model uses the softmax function to calculate the probability distribution of the category to which each node belongs.

2. According to the social robot detection method based on the KAN graph neural network model of claim 1, it is characterized in that: The data extracted in step 1 specifically includes: (1) Node features: including feature tensors extracted from descriptions, tweet-related features, numerical attribute features, which represent the numerical features of nodes, such as the number of followers and tweets of a user, and category attribute features, which represent the category features of nodes; (2) Edge index: used to describe the connection relationship between nodes and indicate the interaction between nodes; (3) Edge type: represents different types of edge relationships, including: follow type, friend type, mention type, reply type, reference type, link, and topic tag; (4) Node label: 0 represents a normal user and 1 represents a social robot account.

3. According to claim 2, a social robot detection method based on a KAN graph neural network model is characterized in that: D min represents the sample set of the minority class, D maj Represents the sample set of the majority class, including N min and N maj samples, for each minority class sample x i ∈D min , use the k nearest neighbor algorithm to find its nearest k neighbors, the neighbors are from the minority class, synthesize more training data for the minority class samples that are more difficult to learn, use the Euclidean distance metric, the number of synthetic samples generated for each minority class sample is based on the complexity of its neighbors, and define the total number of synthetic samples G that need to be generated as: G=β*(N maj -N min ) Among them, β∈[0,1] is a defined balancing factor that controls the number of synthetic samples generated. Then, for each minority class sample x i The number of samples generated g i According to the formula: Among them, n i is the number of neighbors, for each minority class sample x i , according to the calculated g i , randomly select a neighbor x from its neighbors j , generate new synthetic samples by linear interpolation: x new =x i +λ*(x j -x i ) Among them, λ∈[0,1] is a random number used to randomly generate new samples. This formula represents the generated synthetic sample x new Between x i and neighbor samples x j After that, the generated synthetic sample x new Added to the minority class sample set, we finally get a balanced social robot dataset.

4. According to claim 1, a social robot detection method based on a KAN graph neural network model is characterized in that: The model in step 2 combines the advantages of graph convolutional neural networks and KAN neural networks. It can process complex graph structure data and effectively capture nonlinear relationships. One of its core is the graph convolution layer, which is used to aggregate the features of each node and its neighboring nodes and generate new node feature representations.

5. According to the social robot detection method based on the KAN graph neural network model of claim 4, it is characterized in that: Input the data preprocessed in step 1 into the invention model; After two convolutional layers, the core calculation formula is: Among them, H (l) is the node feature representation of the lth layer, A is the original adjacency matrix, D is the degree matrix, is the adjacency matrix after adding self-loops, W (l) is the trainable weight matrix of the lth layer, σ is the nonlinear activation function. This step normalizes the adjacency matrix A, performs weighted aggregation of the features of the node and its neighbors in each layer, and updates the representation of the node.

6. According to the social robot detection method based on the KAN graph neural network model of claim 1, it is characterized in that: In step 3, the input features are transformed nonlinearly through the KAN linear layer to generate new features after B-spline interpolation. The output of the KAN linear layer is then used as part of the node features. Next, these new features are fused with the adjacency matrix features for message passing and aggregation. In this way, each layer of GCN will combine the output of the KAN layer to update the representation of the node.

7. A social robot detection method based on a KAN graph neural network model according to claim 6, characterized in that: The KAN linear layer combines two parts: the basic part based on linear transformation and the nonlinear part based on B-spline interpolation. The KAN linear layer enhances the nonlinear mapping ability of features through B-spline interpolation. Its main function is to introduce a more flexible nonlinear transformation to replace the traditional linear mapping. B-spline interpolation is to calculate the B-spline basis function on a specific grid. The KAN linear layer realizes the local nonlinear interpolation of the input features, so that the features of each node can better adapt to complex relationships. Its main formula is: Among them, f(x1,...,x n ) represents an n-dimensional continuous function, representing the input multidimensional data, Φ q represents a set of trainable one-dimensional functions for nonlinear transformation of the one-dimensional weighted sum of inputs, φ q,p (x p ) is another set of trainable one-dimensional functions that transform each input dimension x p transform; this formula shows that a multivariate function can be fundamentally reduced to a combination of appropriately defined univariate functions, where the combination involves only simple addition.

8. The social robot detection method based on the KAN graph neural network model according to claim 1 is characterized in that: In step 4, in order to prevent the model from overfitting during the training process, the Dropout mechanism is used to randomly inactivate the node features after each layer of the model is output. After each layer of the model is output, the Dropout mechanism is used to randomly inactivate the features to avoid overfitting of the model. After the output of each layer, the Dropout mechanism is used to randomly inactivate the features to avoid overfitting of the model: H (l+1) =Dropout(H (l+1) ,p) Among them, p is the probability of Dropout, which means randomly discarding some features.

9. The social robot detection method based on the KAN graph neural network model according to claim 1 is characterized in that: In step 5, in order to perform the classification task, the model uses the softmax function to calculate the probability distribution of the category to which each node belongs, and the nodes are finally classified by the softmax function.

10. A social robot detection method based on a KAN graph neural network model according to claim 9, characterized in that: After multi-layer graph convolution and feature processing of KAN linear layers, each node will eventually generate a vector representing the node's features. In order to perform classification tasks, the model uses the softmax function to calculate the probability distribution of the category to which each node belongs. Finally, classification is performed through the softmax function: For the GCN classifier, the formula is as follows: in, is the output of the model, i.e., the probability distribution of the predicted label of each node; A is the adjacency matrix, which represents the connection relationship between nodes in the graph; H is the node feature matrix; softmax is the activation function, which converts the output into a probability distribution; To simplify the calculation, the present invention uses the unit matrix E as the input adjacency matrix A; the label and enhancement embedding matrix Input into the formula for training; The formula is as follows: It reduces the complexity of implementation and the difficulty of model design and implementation, making the model more efficient and easier to deploy. It also improves the training and prediction speed of the model, allowing the model to be applied to actual scenarios more quickly. The final objective function of the classifier of the present invention can be written as: Among them, L(y') is the total loss function, y i is node v i The true label, y' i is the corresponding predicted label; V L is the set of nodes containing the true labels, V S is a set containing synthetic nodes; λ is a weight parameter that balances the loss of real nodes and synthetic nodes, ranging from 0 to 1; the loss function measures the supervision loss between the real label and the predicted label.

Images