A social bot detection method based on heterogeneous edge decoupling and balanced label expansion
By constructing a method for heterogeneous edge decoupling and balanced label expansion in social robot detection, the problems of heterogeneous connection influence and pseudo-label expansion bias in social relationship graph representation learning are solved, and more accurate social robot detection and model training results are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SOUTHWEST PETROLEUM UNIV
- Filing Date
- 2026-06-06
- Publication Date
- 2026-08-04
AI Technical Summary
Existing social robot detection methods fail to adequately characterize the impact of heterogeneous connections during the social relationship graph representation learning process, rely on limited labeled samples, and are prone to class bias due to pseudo-label expansion under imbalanced class distribution conditions, affecting the model's discriminative ability and learning performance.
By constructing a method based on heterogeneous edge decoupling and balanced label expansion, edge difference is calculated and consistency propagation weight and difference propagation weight are determined. Pseudo-label samples are screened by combining the prediction consistency of attribute discrimination branch and graph branch, forming a class-balanced expanded training set, reducing heterogeneous connection interference and improving model training effect.
It effectively distinguishes between social bots and real users, reduces the interference of heterogeneous connections on node representation learning, and improves the detection accuracy and training effect of the model under the condition of imbalanced class distribution.
Smart Images
Figure CN122335283B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of social robot detection technology, and in particular to a social robot detection method based on heterogeneous edge decoupling and balanced label expansion. Background Technology
[0002] Social bots pose a persistent security threat to social networking platforms. They can participate in information dissemination and user interaction by automating account control, generating or forwarding content in bulk, and simulating real user interactions, thereby affecting the authenticity and credibility of the social platform ecosystem. As social network structures and user behavior patterns become increasingly complex, accurately identifying social bots has become a crucial issue in network information security and social media governance.
[0003] In recent years, the development of deep learning and Graph Neural Networks (GNN) technologies has enabled social bot detection models to extract deep features from account attributes, textual semantics, and social relationship structures, and to enhance node representations through neighborhood aggregation mechanisms on social relationship graphs (graph structures with accounts as nodes and follow / follow relationships as edges). Existing methods, when considering social relationship structures or heterogeneous relationships, typically address these issues from the perspectives of relationship type modeling, relationship weight setting, or neighborhood information aggregation methods. However, existing graph representation learning methods based on neighborhood aggregation often employ a uniform propagation method for adjacent node information, making it difficult to adequately distinguish between connections between nodes of the same category and connections between nodes of different categories. Connections in real social networks are complex; social bot accounts may establish follow relationships with real user accounts, resulting in connections between nodes of different categories in the social relationship graph. If the model does not adequately characterize the differences between the nodes at both ends of the edges in such connections, information from dissimilar nodes may be directly aggregated into the target node representation during neighborhood propagation, leading to noise propagation or node representation confusion, thereby affecting the discriminative ability of social bot detection models.
[0004] Furthermore, social bot detection typically relies on labeled account samples, but obtaining high-quality labeled samples is costly. Existing methods utilize unlabeled account samples through pseudo-labeling or self-training to expand the training data. In these methods, if pseudo-label selection mainly relies on the prediction results of a single model or a single confidence condition, it may be affected by early prediction errors and class imbalances; erroneous pseudo-labels may also accumulate in subsequent training iterations. Simultaneously, the prediction results of unlabeled samples may concentrate on the majority class, further biasing the expanded training set towards the majority class, thus affecting the model's learning performance on different account categories.
[0005] Therefore, how to combine the difference information of the nodes at both ends of the edge to distinguish the propagation effect of different connections in the process of social relationship graph representation learning, reduce the interference of heterogeneous connections on node representation learning, and improve the utilization quality of unlabeled samples under the conditions of limited labeled samples and uneven class distribution, remains a problem to be solved in social robot detection technology. Summary of the Invention
[0006] To address the technical problems of existing social robot detection methods, such as insufficient characterization of the impact of heterogeneous connections during social relationship graph representation learning, strong dependence on limited labeled samples, and easy generation of class bias by pseudo-label expansion under unbalanced class distribution conditions, this invention provides a social robot detection method based on heterogeneous edge decoupling and balanced label expansion.
[0007] This invention provides a social robot detection method based on heterogeneous edge decoupling and balanced label expansion, comprising the following steps: Step 1: Obtain multimodal features and user relationship data of social network accounts, construct a social relationship graph based on the social relationships between users, map and encode the multimodal features respectively and fuse them to obtain a unified initial representation of nodes; Step 2: Construct an attribute discrimination branch, perform initial category discrimination on each node based on the unified initial representation of the nodes, and output the initial category prediction distribution of each node; Step 3: Based on the difference in representation between the nodes at both ends of each edge in the social relationship graph on the unified initial representation of the node and the difference in the initial category prediction distribution output by the attribute discrimination branch, calculate the edge difference degree of the corresponding edge, and determine the consistency propagation weight and difference propagation weight based on the edge difference degree. Step 4: Aggregate the neighborhood node information based on the consistent propagation weight to obtain the consistent propagation component; extract the difference information between the nodes at both ends of the edge based on the difference propagation weight to obtain the difference propagation component; Step 5: Construct a graph representation learning branch, perform graph representation learning on the nodes based on the consistent propagation component and the differential propagation component, obtain the graph representation of each node, and output the category prediction result of the graph branch; Step 6: Based on the consistency and confidence of the predicted categories of the attribute discrimination branch output and the graph branch output, filter out pseudo-label samples that meet the preset confidence conditions from the unlabeled nodes, and select social robot class samples and real user class samples according to the predicted categories to expand the training set in a class balancing manner. Step 7: Train or iteratively update the model based on the expanded training set, and output the prediction result of whether the target account belongs to a social bot or a real user.
[0008] Further, in step 1, multimodal features and user relationship data of social network accounts are obtained, and a social relationship graph G is constructed based on the social relationships between users; let the numerical features, category features, user description text features, and user tweet text features of node i be denoted as follows: After mapping, the corresponding representations are obtained: ; in, , , and These represent the numerical feature mapping function, the category feature mapping function, the user description text feature mapping function, and the user tweet text feature mapping function, respectively; further fusion yields a unified initial representation of the nodes: ;in, This represents the initial representation of node i. This indicates a splicing operation. This represents the fusion mapping function.
[0009] Furthermore, in step 2, a unified initial representation based on nodes is established. Construct an attribute discrimination branch; this branch uses only the node's own features for initial class discrimination, and outputs the initial class prediction distribution for node i: ;in, and For learning parameters, This represents the initial category prediction distribution obtained by node i based on its own multimodal features, and A represents the attribute discrimination branch. In a binary classification detection scenario, Let be a two-dimensional probability distribution vector, whose two components represent the predicted probabilities that node i belongs to the real user class and the social robot class, respectively.
[0010] Furthermore, in step 3, for any edge This invention comprehensively considers the differences in the unified initial representation of the nodes at both ends of an edge and the differences in the output of the attribute discrimination branch, and defines the edge difference degree as: ; in, Representing an edge The edge difference, after normalization, lies in the interval [0,1]. Let represent the representational difference function between node i and node j on the unified initial representation of nodes. This function represents the difference in distribution between node i and node j on the initial class prediction distribution, where α and β are balance coefficients. This represents the normalization function.
[0011] Furthermore, in step 4, the consistent propagation weight and the difference propagation weight are defined based on the edge difference degree as follows: and ;in, Used to measure edges The degree of contribution to consistent information dissemination The larger the value, the more likely it is to be an edge. The greater the contribution to consistent information dissemination, Used to measure edges The degree of contribution to the preservation of differential information. The larger the value, the more likely it is to be an edge. The greater the contribution to the preservation of difference information, the better; for node i, the consistent propagation component is represented as: The differential propagation component is represented as: ; in, Let i represent the set of neighboring nodes. and This is the learnable parameter matrix.
[0012] Furthermore, in step 5, based on the consistent propagation component... and differential propagation components Obtain the graph representation of node i. The graph branch outputs the category prediction distribution of node i: ;in, and These are learnable parameters; Let G represent the category prediction distribution of node i in the graph branch output, and let G represent the graph branch.
[0013] Furthermore, in step 6, the attribute discrimination branch output is also considered. With graph branch output The prediction consistency and confidence level are considered; only when the predicted categories of unlabeled nodes are consistent in both branches and the overall confidence level is not lower than a preset threshold will they be included in the pseudo-label candidate set, denoted as: ; Where U represents the set of unlabeled nodes, and Let represent the predicted categories of node i by the attribute discrimination branch and the graph branch, respectively. This represents the combined confidence level determined by the prediction confidence levels of the attribute discrimination branch and the graph branch. The preset threshold is indicated. At the same time, the pseudo-label candidate samples are divided into social robot candidate samples and real user candidate samples according to the predicted category. A preset number of samples are selected from the two types of candidate samples according to the comprehensive confidence level to obtain a class-balanced extended sample set.
[0014] Furthermore, in step 7, the model is trained or iteratively updated based on the original labeled samples and the class-balanced extended sample set to obtain the trained social robot detection model, and the trained social robot detection model is used to output the prediction result that the target account belongs to a social robot or a real user.
[0015] By adopting the above technical solution, compared with the prior art, the present invention has at least the following beneficial effects: This invention constructs an edge difference degree based on the differences in node representation and predicted distribution between the nodes at both ends of the edge, and determines the consistency propagation weight and difference propagation weight based on the edge difference degree. For connections with low edge difference degree, its contribution to neighborhood information aggregation is increased; for connections with high edge difference degree, its direct smoothing propagation effect is reduced while retaining the difference information of the nodes at both ends of the edge. This reduces the interference of heterogeneous connections on node representation learning during neighborhood propagation, enabling the model to better distinguish the impact of different connections on the social robot detection results.
[0016] This invention further combines the consistency and overall confidence of the predicted categories of the attribute discrimination branch and the graph branch to screen pseudo-label samples from unlabeled nodes, and selects social robot class samples and real user class samples according to the predicted categories to form a class-balanced extended sample set. In this way, while using unlabeled samples to supplement training data, it reduces the risk of pseudo-label expansion process biasing to the majority class, which is beneficial to improving the model training effect under the condition of class imbalance. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments of the present invention are described below. The drawings are used to illustrate the embodiments of the present invention and do not constitute a limitation on the scope of protection of the present invention.
[0018] Figure 1 This is a flowchart illustrating the social robot detection method based on heterogeneous edge decoupling and balanced label expansion provided in this embodiment of the invention. Detailed Implementation
[0019] The specific embodiments of the present invention will now be described with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention.
[0020] Please see Figure 1 This invention discloses a social robot detection method based on heterogeneous edge decoupling and balanced label expansion. In one embodiment, taking the publicly available social robot detection datasets TwiBot-20 and TwiBot-22 as an example, let the social relationship graph be... Where V is the set of user nodes, and E is the set of social relationship edges between users; the set of node features is denoted as X, and the set of category labels for labeled nodes is denoted as Y; in the category labels, social robots are denoted as 1, and real users are denoted as 0; let the total number of nodes in the graph be N, and the set of labeled nodes be denoted as... Let U be the set of unlabeled nodes, then we have .
[0021] Step S101: Obtain multimodal features and user relationship data of social network accounts, construct a social relationship graph based on the social relationships between users, and map, encode and fuse the multimodal features to obtain a unified initial representation of nodes; In this embodiment, for any node i, its input features include four categories: numerical features Category characteristics User description text features and user tweet text features Among them, numerical features are used to characterize the continuous attribute information of user accounts, including at least one or more of the following: number of followers, number of posts, number of favorites, account registration duration, and username length; categorical features are used to characterize the discrete attribute information of user accounts, including at least one or more of the following: whether the account is verified, whether the default avatar is used, whether geolocation is enabled, and whether the account is protected; user description text features are the personal profile text on the user's homepage; and user tweet text features are the text sequence formed by the user's historical tweets. The four types of features are mapped and encoded respectively. Numerical features are normalized and then input into a fully connected layer to obtain numerical feature representations. Categorical features are one-hot encoded, embedded, or processed by a fully connected layer to obtain categorical feature representations. User description text features and user tweet text features are encoded by a text encoder to obtain their corresponding text feature representations. After mapping and encoding, the following results are obtained: ; in, , , and These represent the numerical feature mapping function, the category feature mapping function, the user description text feature mapping function, and the user tweet text feature mapping function, respectively. , , and Let i represent the numerical feature representation, category feature representation, user description text feature representation, and user tweet text feature representation, respectively; further fusion yields a unified initial representation for the node: ; in, This represents the initial representation of node i. This indicates a splicing operation. This represents the fusion mapping function.
[0022] Step S102: Construct an attribute discrimination branch, perform initial category discrimination on each node based on the unified initial representation of the nodes, and output the initial category prediction distribution of each node; In this embodiment, the attribute discrimination branch only uses the node's own multimodal information for category discrimination, and is used to output attribute prior results that do not depend on graph structure propagation; specifically, for node i, the node is uniformly initialized. Input attribute discrimination branch to obtain the initial class prediction distribution of node i: ;in, and Learnable parameters for attribute-based decision branches. Let A represent the initial category prediction distribution obtained by node i based on its own multimodal features, and let A represent the attribute discrimination branch; in a binary classification detection scenario, Let i be a two-dimensional probability distribution vector, whose two components represent the predicted probabilities of node i belonging to the real user class and the social robot class, respectively; during the training phase, the supervised loss of the attribute discrimination branch is defined as: ; in, This represents the set of labeled nodes. This represents the one-hot vector corresponding to the true label of node i; the output of the attribute discrimination branch. In subsequent steps, it is used for edge difference calculation and screening of pseudo-labels for unlabeled nodes. The Softmax function represents the normalized exponential function, and the one-hot vector represents the one-hot encoded vector. In one embodiment, the dimension of the hidden layer of the attribute discrimination branch is set to 128, the activation function is ReLU, and the weight decay coefficient of 0.0001 is used in the optimization process.
[0023] Step S103: Based on the differences between the nodes at both ends of each edge in the unified initial representation and the initial category prediction distribution, calculate the edge difference degree and determine the consistency propagation weight and the difference propagation weight; in this embodiment, for any edge in the graph... The edge difference degree is calculated based on the degree of difference between the two endpoints of the edge in the unified initial representation space and the degree of difference in the output of the attribute discrimination branch. First, define the node representation of the difference item as: ; in, and Let i and j represent the unified initial representations of nodes i and j, respectively. This represents the transpose of a vector. The L2 norm of a vector. This represents the cosine difference between node i and node j in the unified initial representation space of nodes.
[0024] Furthermore, the predicted distribution difference term is defined as: ; in This represents the intermediate distribution obtained by averaging the initial class prediction distributions of nodes i and j. Denotes KL divergence, The edge dissimilarity function represents the distribution difference between node i and node j on the initial class prediction distribution; combining the above two parts, the edge dissimilarity is defined as: ; in, and This is a non-negative balance coefficient used to adjust the contributions of node representation difference terms and predicted distribution difference terms in edge difference degree calculation. This represents the normalization function, which in this embodiment is the Sigmoid function; further, based on the edge difference degree... The consistent propagation weight and the differential propagation weight are determined as follows: and ;in, Representing an edge The degree of contribution to consistent information dissemination Representing an edge The extent of its contribution to the preservation of differential information; The smaller the value, the smaller the difference between the nodes at both ends of the edge. Consistent propagation weight The larger; The larger the value, the greater the difference between the nodes at both ends of the edge. Differential propagation weight The larger.
[0025] Step S104: Aggregate neighborhood node information based on consistent propagation weight to obtain consistent propagation component; extract difference information of nodes at both ends of edge based on difference propagation weight to obtain difference propagation component; In this embodiment, the node unified initial representation obtained in step S101 is used. As the input representation of the 0th layer of the graph convolutional network, and based on the consistent propagation weights and differential propagation weights obtained in step S103, a consistent propagation channel and a differential propagation channel are constructed respectively to perform hierarchical updates on the node representation; for the l-th layer graph convolution, let the representation of node i in the l-th layer be as follows: In the consistent propagation channel, based on the consistent propagation weights between node i and its neighboring nodes, the node representations of the neighboring nodes at layer l are weighted and aggregated to obtain the consistent propagation component of node i at layer l: ;in, This is the normalized result of the consistent propagation weights; correspondingly, in the difference propagation channel, based on the difference propagation weights between node i and its neighboring nodes, the representation differences between node i and its neighboring nodes at layer l are extracted using weighted methods to obtain the difference propagation components of node i at layer l: ;in, This is the normalized result of the difference propagation weights; in this embodiment, the two propagation channels and the node's own residual term are used together to update the next layer node representation: ;in, This represents the residual mapping matrix of the l-th layer. This represents a non-linear activation function; in one embodiment, the number of graph convolutional layers L is 2, and the hidden dimension is 128.
[0026] Step S105: Construct graph representation learning branches. Graph representation learning is performed based on the consistent propagation component and the differential propagation component to obtain the graph representation of each node, and the graph branch category prediction result is output. In this embodiment, after the L-layer graph convolutional propagation in step S104, the final node representation of node i is obtained. And this final node representation is used as the graph representation of node i: ;in, The graph representation of node i, the graph representation It also includes neighborhood information with low edge dissimilarity obtained by aggregating nodes in the consistent propagation channel, as well as the difference information of the nodes at both ends of the edge retained in the difference propagation channel; the graph representation of node i The input graph branch output layer yields the graph branch category prediction distribution for node i: ; in, and These are the weight matrix and bias term of the graph branch output layer, respectively. Let represent the predicted probability that node i belongs to the social bot class or the real user class; in this embodiment, the supervision loss of the graph branch is defined as: ; in, This represents the set of labeled nodes. Let i be the one-hot vector of the true label of node i. This represents the vector transpose. To reduce the prediction bias between the attribute discrimination branch and the graph branch at labeled nodes, a consistency constraint term between the attribute discrimination branch and the graph branch is introduced in this embodiment: ; in, Used for constraint attribute discrimination branch output With graph branch output The predicted distribution remains relatively consistent across labeled nodes.
[0027] Step S106: Based on the consistency and confidence of the predicted categories of the attribute discrimination branch output and the graph branch output, filter pseudo-label samples that meet the preset confidence conditions from the unlabeled nodes, and expand the training set in a balanced manner according to the categories; in this embodiment, step S106 is used to expand the training samples using unlabeled nodes when the number of labeled nodes is limited; let the set of unlabeled nodes be: For any unlabeled node Branches are output based on attributes respectively. And graph branch output Calculate the prediction category and prediction confidence.
[0028] Specifically, for any unlabeled node The attribute-based branch prediction category of node i is denoted as: The branch prediction category is denoted as Correspondingly, the maximum prediction probabilities of the attribute discrimination branch and the graph branch are denoted as follows: and In this embodiment, the average of the prediction confidence scores of the two branches is used as the overall confidence score of node i. ;in, The overall confidence level of node i is represented by: An unlabeled node is included in the pseudo-label candidate set only when the predicted categories of the unlabeled node are consistent in both the attribute discrimination branch and the graph branch, and the overall confidence level is not lower than a preset threshold. The pseudo-label candidate set is represented as: ; in, This is the confidence threshold. In one embodiment, Taking 0.8, we obtain the candidate set of pseudo-labels. Subsequently, to prevent the pseudo-label expansion process from continuously biasing towards the majority class, the candidate set was divided into a social robot candidate set and a real user candidate set, denoted as follows: and Then, according to the overall confidence level... The top M social bot samples and the top M real user samples are selected from the candidate sets of social bots and real users to form a balanced expanded sample set. To ensure that the expanded training samples remain relatively balanced across categories; when the number of candidate samples in any category is less than the preset value M, the smaller of the number of candidate samples in the two categories is used as the actual selection number, and the same number of samples are selected from the social robot candidate set and the real user candidate set respectively.
[0029] In this embodiment, pseudo-label expansion is performed iteratively; firstly, the original set of labeled nodes is used. The model is pre-trained until a preset number of warm-up rounds are reached. Then, pseudo-label filtering and class balancing expansion are initiated; thereafter, after a preset training interval... Re-evaluate attribute discrimination branches and graph branch predictions for unlabeled nodes, and update the pseudo-label candidate set. and a class-balanced extended sample set In one embodiment, a preset number of preheating cycles is used. Set the value to 20, and set the training interval to 20. Take 5.
[0030] Step S107: Train or iteratively update the model based on the expanded training set, and output the prediction result that the target account belongs to a social robot or a real user. In this embodiment, model training includes a supervised pre-training phase and a balanced label expansion training phase; in the supervised pre-training phase, based on the set of labeled nodes... The training attributes discriminative branch and the graph branch are used, and the model's total loss function is defined as follows: ; in, This represents the total loss during the supervised pre-training phase. This represents the supervised loss for the attribute discrimination branch. This represents the supervised loss of the graph branch. This represents the consistency constraint term between attribute-determining branches and graph branches. This is the balance coefficient. In one embodiment, it is taken as... During the balanced label expansion training phase, the class-balanced expanded sample set is used. Introducing the training process; for the attribute discrimination branch, its pseudo-label loss is defined as: ; For the graph branch, the pseudo-label loss is defined as: ; in, The pseudo-label loss represents the attribute discrimination branch. The pseudo-label loss represents the graph branch; further, the total loss function for the balanced label expansion training phase is defined as: ; in, This represents the pseudo-label loss weight coefficient, used to control the degree of influence of pseudo-labeled samples on model training; in one embodiment... The model is trained using the Adam optimizer with an initial learning rate of 0.001 and a total of 200 training epochs. If the validation set metric does not improve for 20 consecutive epochs, training is terminated using an early stopping strategy. After training, the multimodal features and user relationship data of the target account are input into the trained social bot detection model to obtain the predicted probabilities of the target account belonging to the social bot or real user category. Based on the predicted probabilities, the model outputs the prediction result of whether the target account belongs to a social bot or a real user.
[0031] The specific embodiments of the present invention have been described above with reference to the accompanying drawings, but the present invention is not limited to the above embodiments.
[0032] It should be noted that, for those skilled in the art, without departing from the concept of the present invention, adjustments or substitutions can be made to the specific calculation method of edge difference degree, the specific construction method of consistent propagation component and difference propagation component, the specific implementation method of graph representation learning module, the pseudo-label filtering rules and the specific parameter settings of balanced label expansion; the above adjustments or substitutions, without departing from the essence of the technical solution of the present invention, should all be regarded as equivalent changes or improvements of the present invention.
[0033] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the scope of protection of the present invention. Any modifications or equivalent substitutions made by those skilled in the art to the foregoing embodiments without departing from the essence of the technical solutions of the present invention should fall within the scope of protection of the present invention.
Claims
1. A social bot detection method based on decoupling with heterogeneous edge and equalizing label expansion, characterized in that, Includes the following steps: Step 1: Obtain multimodal features and user relationship data of social network accounts, construct a social relationship graph based on the social relationships between users, map and encode the multimodal features respectively and fuse them to obtain a unified initial representation of nodes; Step 2: Construct an attribute discrimination branch, perform initial category discrimination on each node based on the unified initial representation of the nodes, and output the initial category prediction distribution of each node; Step 3: Based on the difference in representation of the nodes at both ends of each edge in the social relationship graph on the unified initial representation of the node and the difference in the initial category prediction distribution output by the attribute discrimination branch, calculate the edge difference degree of the corresponding edge, and determine the consistency propagation weight and difference propagation weight based on the edge difference degree. Step 4: Aggregate the neighborhood node information based on the consistent propagation weight to obtain the consistent propagation component; extract the difference information between the nodes at both ends of the edge based on the difference propagation weight to obtain the difference propagation component. Step 5: Construct graph representation learning branches, perform graph representation learning on nodes based on the consistent propagation component and the differential propagation component, obtain the graph representation of each node, and output the category prediction results of the graph branches; Step 6: Based on the consistency and confidence of the predicted categories of the attribute discrimination branch output and the graph branch output, filter out pseudo-label samples that meet the preset confidence conditions from the unlabeled nodes, and select social robot class samples and real user class samples according to the predicted categories to expand the training set in a class balancing manner. Step 7: Train or iteratively update the model based on the expanded training set, and output the prediction result of whether the target account belongs to a social bot or a real user; In step 3, for any edge in the social relationship graph... The edge dissimilarity degree is calculated based on the differences in the initial representations of the nodes at both ends of the edge and the differences in the initial category prediction distributions of the attribute discrimination branch outputs. ;in, Representing an edge The edge difference, after normalization, lies in the interval [0,1]. Let represent the representational difference function between node i and node j on the unified initial representation of nodes. This function represents the difference in distribution between node i and node j on the initial class prediction distribution, where α and β are balance coefficients. Represents the normalization function; in step 3, the consistency propagation weight and difference propagation weight are determined based on the edge difference degree: and ;in, Used to measure edges The degree of contribution to consistent information dissemination The larger the value, the more likely it is to be an edge. The more suitable it is for consistent information dissemination, Used to measure edges The degree of contribution to the preservation of differential information. The larger the value, the more likely it is to be an edge. The more important it is to preserve the differences between the nodes at both ends of an edge; In step 4, for node i, the information of neighboring nodes is aggregated based on the consensus propagation weight to obtain the consensus propagation component. Based on the difference propagation weights, the difference information between the nodes at both ends of the edge is extracted to obtain the difference propagation components. ;in, Let i represent the set of neighboring nodes. and For learnable parameter matrix, This represents the consistent propagation component of node i. This represents the differential propagation component of node i.
2. The social robot detection method based on heterogeneous edge decoupling and balanced label expansion according to claim 1, characterized in that, In step 1, the multimodal features include numerical features, categorical features, user description text features, and user tweet text features; let the numerical features, categorical features, user description text features, and user tweet text features of node i be denoted as follows: After mapping and encoding, the corresponding representations are obtained: ； in, , , and These represent the numerical feature mapping function, the category feature mapping function, the user description text feature mapping function, and the user tweet text feature mapping function, respectively. , , and Let i represent the numerical feature representation, category feature representation, user description text feature representation, and user tweet text feature representation, respectively; further fusion yields a unified initial representation for the node: ;in, This represents the initial representation of node i. This indicates a splicing operation. This represents the fusion mapping function.
3. The social robot detection method based on heterogeneous edge decoupling and balanced label expansion according to claim 1, characterized in that, In step 2, the attribute discrimination branch only uses the node's own multimodal features for initial category discrimination. For node i, the node's unified initial representation is input into the attribute discrimination branch to obtain the initial category prediction distribution of node i: ;in, and Learnable parameters for attribute-based decision branches. Let A represent the initial category prediction distribution obtained by node i based on its own multimodal features, and let A represent the attribute discrimination branch; in a binary classification detection scenario, Let be a two-dimensional probability distribution vector, whose two components represent the predicted probabilities that node i belongs to the real user class and the social robot class, respectively.
4. The social robot detection method based on heterogeneous edge decoupling and balanced label expansion according to claim 1, characterized in that, In step 5, based on the consistent propagation component and differential propagation components Obtain the graph representation of node i. The graph branch outputs the category prediction distribution of node i: ;in, and For learnable parameters, Let G represent the category prediction distribution of node i in the graph branch output, and let G represent the graph branch.
5. The social robot detection method based on heterogeneous edge decoupling and balanced label expansion according to claim 1, characterized in that, In step 6, the branch output is based on attribute discrimination. And graph branch output Determine the prediction category and overall confidence level of the unlabeled nodes; An unlabeled node is included in the pseudo-label candidate set only if its predicted class is consistent in both branches and its overall confidence score is not lower than a preset threshold. ; Where U represents the set of unlabeled nodes, and Let represent the predicted categories of node i by the attribute discrimination branch and the graph branch, respectively. This represents the combined confidence level determined by the prediction confidence levels of the attribute discrimination branch and the graph branch. This indicates a preset threshold.
6. The social robot detection method based on heterogeneous edge decoupling and balanced label expansion according to claim 5, characterized in that, In step 6, the samples in the pseudo-label candidate set are divided into social robot candidate samples and real user candidate samples based on the predicted category. A preset number of samples are selected from the social robot candidate samples and the real user candidate samples according to the comprehensive confidence level to obtain a class-balanced extended sample set.
7. The social robot detection method based on heterogeneous edge decoupling and balanced label expansion according to claim 6, characterized in that, In step 7, the model is first pre-trained based on the original labeled samples. After reaching the preset number of warm-up rounds, the class-balanced extended sample set is added to the training process, and the unlabeled nodes are re-predicted using a two-branch method according to the preset training interval to update the pseudo-label candidate set and the class-balanced extended sample set. After training is completed, the prediction result of whether the target account belongs to a social robot or a real user is output.