A social robot detection method based on multi-view graph representation learning

Abstract

This invention discloses a social robot detection method based on multi-view graph representation learning, comprising the following steps: Step 1, acquiring user data from a social platform to construct a heterogeneous graph; Step 2, encoding multimodal attributes for feature initialization; Step 3, constructing a community-aware view, a noise-aware view, and a motif-aware view, and obtaining node representations under different views; Step 4, adaptively fusing node features from different views for classification prediction, and using contrastive learning as an auxiliary task to optimize node representations: a gated attention mechanism is adopted, combined with classification loss, contrastive loss, and diffusion reconstruction loss for joint training. This invention can effectively address the problem of low detection accuracy in existing technologies when facing structural heterogeneity, adversarial relationship masquerading, and group collaborative behavior of social robots, thereby achieving accurate identification of social robots.

Description

Technical Field

[0001] This invention relates to the field of social network technology, and more specifically to a social robot detection method based on multi-view graph representation learning. Background Technology

[0002] Online social media platforms, exemplified by Twitter and Facebook, have evolved into crucial channels for information dissemination, opinion exchange, and public communication. However, the virtual nature of these platforms provides fertile ground for the proliferation of social bots. Social bots are accounts that are automatically controlled by programs and capable of simulating the behavior of real users. Once maliciously deployed, these accounts can spread disinformation on a large scale within a short period, manipulate public opinion, create social division, and even interfere with political activities and public decision-making processes, severely distorting public perception and undermining the credibility and security of cyberspace. Therefore, accurately detecting social bots in the complex environment of social networks has become an urgent task for maintaining cyberspace security.

[0003] Early social bot detection methods primarily relied on extracting manually generated features from user metadata to train classifiers. These included extracting account attributes, text content features, and behavioral statistics from user metadata, and combining this with traditional machine learning models for discrimination. However, with the continuous development of social bot technology, their behavioral patterns have become increasingly complex, and their simulations of human behavior have become more realistic. Simple manually generated features alone are no longer sufficient to effectively identify highly sophisticated bots with excellent disguises. In recent years, researchers have begun to utilize graph neural networks to construct detection methods from the perspective of network topology. These methods typically model social networks as graphs, where users are abstracted as nodes, and user interactions (such as following, mentioning, and forwarding) are represented as edges. Compared to traditional methods, graph-based methods have achieved significant progress in detection performance by jointly modeling graph structure information and account-level metadata.

[0004] Although existing graph-based social robot detection technologies have made substantial progress, they still have the following shortcomings when facing complex, elusive, and adversarial advanced social robots:

[0005] (1) Ignoring the hierarchical organizational structure of social networks. Users in social networks are usually clustered into hierarchical communities of different granularities. However, existing technologies usually divide the network into disjoint or flat communities, making it difficult to characterize the inherent hierarchical semantics. In real-world scenarios, social bots often exist covertly as small-scale, high-density subgroups and are embedded in larger-scale human-machine hybrid communities. Single-scale structural modeling methods are difficult to identify such anomalous groups from a global perspective.

[0006] (2) Lack of ability to identify fake relationships. Social bots often disguise their identities by actively interacting with real users, rather than simply establishing connections with similar accounts. This behavior undermines the homogeneity assumption in social networks and introduces a large number of noisy edges into the observation graph structure. Existing technologies usually assume that all observation edges have equal reliability, and cannot distinguish between real and fake relationships during feature aggregation. This can easily lead to excessive mixing of robot and human node features, resulting in oversmoothing and severely limiting the model's discriminative ability.

[0007] (3) Difficulty in capturing high-order collaborative behaviors. Social bots typically perform specific tasks in a collaborative manner, such as collaborative forwarding, concentrated attacks, or public opinion manipulation. These collaborative behaviors are often manifested in specific high-order structural patterns in network structures, such as star topologies, ternary closures, or high-density connection motifs. However, most existing methods focus on pairwise relationships between nodes and lack the ability to effectively model the semantics of high-order structures, making it difficult to identify highly identifiable groups of collaborative bots.

[0008] In summary, existing technologies have significant limitations in handling the structural heterogeneity, adversarial relationship masquerading, and high-order collaborative behaviors of social robots. A key challenge in the field of social robot detection is how to construct a robust multi-view framework from multiple perspectives, including global community analysis, local denoising, and high-order semantics, to uncover deep-seated anomaly patterns in social robots. Summary of the Invention

[0009] The purpose of this invention is to provide a social robot detection method based on multi-view graph representation learning, which solves the problems of existing technologies in detecting social robots, such as ignoring the hierarchical community organizational structure of social networks, difficulty in suppressing structural noise introduced by adversarial relationship masquerading, and inability to effectively model high-order collaborative behavior patterns, thereby improving the detection accuracy and robustness in complex adversarial environments.

[0010] To achieve the above objectives, the present invention adopts the following technical solution: A social robot detection method based on multi-view graph representation learning includes the following steps: Step 1: Obtain user data from social platforms and construct a heterogeneous graph: Extract user metadata and social relationship data from social platforms and perform data preprocessing operations to model the online social network as a heterogeneous graph. Assign category labels of social robots or real users to some user nodes based on manual annotation. Step 2, Encoding Multimodal Attributes for Feature Initialization: The preprocessed user metadata is divided into three categories: numerical attributes, boolean attributes, and text content. Initial feature vectors are generated through parametric feature transformation and processing by the pre-trained language model RoBERTa. And construct a directed social graph based on attention relationships. ,in, It is a set of nodes, where each node represents a user of a social media platform. To describe the edge set of user attention relationships, and thus based on and Train the CNMBot model to detect whether a user is a social bot based on their social relationships and attribute features; Step 3: Construct a community-aware view, a noise-aware view, and a motif-aware view, and obtain node representations under different views: based on the directed social graph. and initial feature vector From the perspectives of community discovery, relationship denoising, and behavioral collaboration, a community perception view, a noise perception view, and a motif perception view are constructed respectively, thereby obtaining the node representation under the community perception view. Node representation under the relationship denoising view Node representation under Motif-aware view To characterize the hierarchical community organizational structure among users, the reliability of interaction relationships, and the high-level collaborative behavior characteristics of user groups; Step 4: Adaptively fuse node features from different views for classification prediction, and use contrastive learning as an auxiliary task to optimize node representation: Employ a gated attention mechanism to calculate the gate signal for each view. By adaptively fusing node features from different views through element-wise multiplication, a fused comprehensive representation is obtained. ,Will The input is a multilayer perceptron for classification prediction, and the output is the prediction result of the corresponding node as a social robot or a human user. An auxiliary contrastive learning task with the noisy perception view as the anchor point is introduced, and the classification loss, contrastive loss and diffusion reconstruction loss are combined for joint training.

[0011] Optionally, step 1 includes modeling the online social network as a heterogeneous graph structure. The task of detecting social bots on social platforms is described as a binary classification problem of graph nodes, where It is a set of user nodes, each node representing a social media account, with a total of [number missing] nodes. ; To describe the edge set of interactions between users, It is a set of relationship types, which includes at least the user's follow and follow-up relationships; It is the metadata matrix of the nodes, where Representing the Metadata for each node, including the user's numerical statistical characteristics. Boolean properties and text content , Represents the set of node labels, where, Instructing users Is it a bot account? This indicates the account is a bot, 0 represents a human, and the ultimate goal is to learn a mapping function. This allows each node to be categorized based on social network relationships and user attributes. Mapped to its corresponding category label.

[0012] Optionally, step 2 specifically includes: Step 21, for each user Considering the properties of the three modalities, namely numerical statistical characteristics Boolean properties and text content In this process, the text content is encoded using a pre-trained language model, RoBERTa, to capture semantic information, while numerical and categorical features are preprocessed using a learnable projection matrix. and Each feature is converted into a vector through linear mapping, and then concatenated with the three types of feature representations mentioned above before being input into a multilayer perceptron for nonlinear fusion processing to obtain the nodes. initial features , Formula 1; in, and It is a projection matrix. Represents the language model encoder, Indicates splicing operation Step 22, in the relationship graph construction section, a directed graph is constructed based on the follow and follow relationships between users. The constructed graph will be compared with the initial feature matrix Together, they serve as inputs for subsequent multi-view learning modules to construct the CNMBot model.

[0013] Optionally, step 3 specifically includes: Step 31: Construct a community-aware view to capture the global hierarchical community structure, identify potential social robot clusters obscured by structural noise and dummy connections, and learn node representations. ; Step 32: Construct a noise-aware view by modeling edge reliability through a generative diffusion process to filter out unreliable and spoofed relationships, and learn node representations. ; Step 33, constructing the Motif-aware view, involves explicitly modeling the collaborative interaction patterns of the social robot by mining higher-order patterns, and learning node representations. .

[0014] Optionally, step 31 specifically includes: Step 311, specify the tree height as By minimizing the social graph structural entropy To construct a hierarchical coding tree ; Step 312: Utilize a structural entropy pooling network to aggregate node features layer by layer along the encoding tree, capturing global hierarchical community patterns masked by structural noise, and obtaining node representations under the community-aware view. , Formula 2; in, This represents the hierarchical aggregation function of SEP-N. These are the parameters of the structural entropy pooling network. The initial node features are used; the mechanism gradually abstracts the features from the local neighborhood to the global community context, generating a high-order representation that is robust to structural perturbations.

[0015] Optionally, step 32 specifically includes: Step 321: Using a denoised diffusion probability model, the observed social edges are treated as latent variables containing noise. The forward process is in Gaussian noise is gradually injected into the edge variables within each time step, converting the discrete edge indicators into continuous noise observations. The reverse process involves introducing a learnable denoiser. Embedded by edges and time step embedding To conditionally predict noise, minimize the difference between the predicted noise and the actual noise. Optimize the denoiser based on the mean square error between them ; Step 322, from the time step pure Gaussian noise distribution We begin by reconstructing edges using the learned denoiser and then applying sigmoid activation to calculate confidence scores. A weighted mask is generated using confidence scores. ; Step 323: Aggregate information on reliable social relationships through a relational graph convolutional network to obtain node representations under a noise-aware view. , Formula 3; in, Indicates by parameters The message passing function is driven by this diffusion denoising mechanism, which allows the encoder to focus more on aggregating information from reliable social relationships by evaluating the reliability of the edges.

[0016] Optionally, step 33 specifically includes: Step 331 constructs three semantic adjacency matrices corresponding to the basic structure Motif. One of the first-order connections is a Motif Capturing star-shaped structure and fundamental connectivity, constructed using an undirected adjacency matrix: Two-way interactive Motif By identifying bidirectional edges to model mutual social trust, and constructing the Hadamard product: ; Ternary closure Motif Capturing community cohesion and coordinated interactions through the use of binary matrices Extracting triangular structures from user interactions: ; Step 332: Employ a multi-channel graph convolutional network to aggregate information within these semantic spaces, for each motif The node feature update is shown in Equation 4. Equation 4; Step 333: Apply an attention mechanism to aggregate the embeddings learned from the three channels to obtain the node representations under the final Motif-aware view. , Formula 5; in, This represents the Motif aggregation function, which combines Motif convolution with attention fusion, and is defined by parameters. The mechanism learns higher-order representations by modeling collaborative behaviors among users as motifs.

[0017] Optionally, step 4 specifically includes: Step 41: A gating attention mechanism is designed to adaptively fuse node features from three views to obtain the fused node representation. , (Right now The (line) is the user Feature representation after fusing different views; Step 42, Node Representation The data is fed into a multilayer perceptron to predict the probability that each user is a social bot; for each node... , Represents the model for nodes The predicted probability of belonging to the category, Formula 6; The classifier is trained in a supervised manner using binary cross-entropy loss, and the classification loss is... Calculated using Equation 7 Formula 7; in, Represents the set of training nodes. For nodes The true label; Step 43 introduces self-supervised contrastive learning as an auxiliary task to align the semantics between different views and enhance the robustness of node representations; noise-aware views are then used. When specified as an anchor view, the goal of contrastive learning is to maximize the anchor view. With two complementary views and Mutual information between them; two alignment items and Calculated using Equations 8 and 9 Formula 8; Equation 9; Among them, losses Use InfoNCE loss; Step 44: Perform multi-objective joint optimization to calculate the final loss of CNMBot. It is a weighted combination of classification, comparison, and diffusion reconstruction losses. Formula 10; in, , , These are hyperparameters for balancing auxiliary objectives. For binary cross-entropy classification loss, and The InfoNCE loss represents the relationship between the representations learned by the community-aware view, the motif-aware view, and the noise-aware view, respectively. This represents the reconstruction mean square error loss for the diffusion model.

[0018] Optionally, step 41 specifically includes: Step 411, for each view , where the view Represents a community-aware view, view Represents a noise-perceived view, view Representing the Motif-sensing view, the gating signal is calculated using Equation 11. , Formula 11; in, This represents the gating signal vector corresponding to view k, used to adaptively measure the contribution of that view to node identity determination. This represents the learnable weight matrix corresponding to view k. This represents the learnable bias vector corresponding to view k. For activation functions; Step 412: Element-wise multiply the features learned from each view with the gating signal to obtain the fused node representation. , Equation 12; in, This represents the Hadamard product.

[0019] Compared with the prior art, the present invention has the following advantages: 1. The noise perception view designed in this invention uses a generative diffusion denoising mechanism to model the reliability of edges in social networks as a latent variable, thereby actively identifying and weakening the structural noise introduced by the fake relationships deliberately constructed by social robots. This effectively avoids the indiscriminate feature propagation between robot nodes and human nodes, suppresses the phenomenon of excessive feature aggregation, and enhances the robustness of detection.

[0020] 2. The community-aware view constructed in this invention is based on the principle of minimizing structural entropy. It systematically decodes the inherent hierarchical organization formed by coarse-grained groups to fine-grained sub-communities in social networks, which can effectively reveal the semantic relationships of multi-level nested communities, thereby identifying robots hidden in complex human-machine hybrid groups.

[0021] 3. The Motif perception view introduced in this invention explicitly models high-order structures with discriminative value, such as ternary closures, bidirectional interactions, and star topologies, thereby characterizing the group collaboration mode of social robots when performing tasks and capturing the complex collaborative manipulation behavior among social robots.

[0022] This invention further designs an adaptive gating fusion mechanism, which dynamically adjusts the weight allocation of different view representations during the fusion process based on the importance of node features, and combines a contrastive learning strategy to enhance the consistency and complementarity among multiple view representations. Through this fusion method, the model can fully leverage the advantages of each view, making the inter-class boundaries between social robots and real users in the representation space clearer. Attached Figure Description

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the following detailed description to explain the invention, but do not constitute a limitation thereof. In the drawings: Figure 1 A schematic diagram of the process for the social robot detection method based on multi-view graph representation learning provided by the present invention; Figure 2 This is a diagram illustrating the overall framework of the social robot detection method based on multi-view graph representation learning provided by the present invention. Figure 3 To construct a coding tree by minimizing structural entropy (in terms of height) (For example) and a schematic diagram of the principle of the corresponding cluster allocation matrix; Figure 4 The experimental comparison graph shows the impact of different negative sampling rates on various performance metrics of the method of this invention on the TwiBot-20 and MGTAB datasets; Figure 5 The experimental comparison graph shows the impact of different diffusion steps on various performance metrics of the method of this invention on the TwiBot-20 and MGTAB datasets; Figure 6 Experimental comparison of the impact of different coding tree depths on various performance metrics of the method of this invention on the TwiBot-20 and MGTAB datasets; Figure 7 For different contrast loss coefficients and A heatmap showing the change in detection accuracy of the method of the present invention under combined conditions; Figure 8 For different diffusion loss coefficients Experimental comparison of the impact of the method of the present invention on various performance indicators on the TwiBot-20 and MGTAB datasets; Figure 9 This is a visualization comparing the clustering results of node representations learned by different methods on the TwiBot-20 dataset after processing with t-SNE. Detailed Implementation

[0024] The technical solution of the social robot detection method based on multi-view graph representation learning proposed in this invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.

[0025] Social bots are user accounts controlled by automated programs. They are widely present on social media platforms such as Twitter, Facebook, TikTok, and Weibo, and can simulate real user behavior by posting content, establishing connections, and participating in interactions. In application scenarios such as information recommendation, public opinion analysis, content moderation, and public safety, if social bots are used to spread false information, manipulate public opinion, or amplify specific viewpoints abnormally, they can easily interfere with the platform's operational order and the credibility of information. Therefore, there is an urgent need for effective identification and detection of social bots.

[0026] As social bots become increasingly adept at masquerading their behavior patterns and relationship structures, they often proactively construct social relationships similar to those of real users and participate in group interactions. This renders traditional detection methods based on single features or local relationships insufficiently robust in complex social network environments. To address this, this invention proposes a social bot detection method based on multi-view graph representation learning. By modeling online social networks as user interaction graphs, this method models these graphs from three complementary perspectives: community structure, reliability of interaction relationships, and collaborative behavior of user groups. This comprehensive approach aims to characterize the differences in organizational structure, relationship composition, and behavioral patterns of social bots, thus providing a technical reference for automated account identification on social platforms.

[0027] This invention presents a social bot detection method based on multi-view graph representation learning, applicable to scenarios such as malicious automated account identification, online social network security governance, online public opinion monitoring and guidance, and prevention and control of false information dissemination. The proposed method encodes the multimodal information of social platform users and constructs a social graph based on user attention relationships to obtain initial node features and graph structure. It then designs three views—community perception, noise perception, and motif perception—from three perspectives: user community affiliation, relationship reliability, and collaborative behavior patterns. The community perception view utilizes structural entropy minimization to reveal hierarchical community patterns hidden in the global topology; the noise perception view introduces a generative diffusion mechanism to identify and suppress disguised noise edges used by social bots to evade detection; and the motif perception view aggregates high-order structural motifs to model collaborative behavior. Furthermore, a gated attention mechanism is introduced to adaptively fuse feature representations learned from different views, and auxiliary contrastive learning is combined to enhance the robustness of the representations. This invention can effectively address the problem of low detection accuracy of existing technologies when facing structural heterogeneity of social robots, adversarial relationship masquerading, and group collaborative behavior, so as to achieve accurate identification of social robots.

[0028] The present invention provides a social robot detection method based on multi-view graph representation learning, comprising: acquiring user data from a social platform to construct a heterogeneous graph; encoding multimodal information of users in the social network; constructing initial feature representations of nodes and user interaction graph structure; modeling the user interaction graph from three perspectives: community discovery, relationship denoising, and behavioral collaboration, respectively characterizing the hierarchical community organizational structure among users, the reliability of interaction relationships, and the high-order collaborative behavior features of user groups; obtaining node representations under multiple views; adaptively fusing heterogeneous node representations from different views based on a gating attention mechanism to obtain a unified node representation; inputting the unified node representation into a classification prediction model; outputting a prediction result that the corresponding node is a social robot or a real user; and combining a contrastive learning strategy to align the multi-view node representations to improve the discriminativeness and stability of classification.

[0029] Specifically, the following steps are included: Step 1: Obtain user data from social platforms and construct a heterogeneous graph: Extract user metadata and social relationship data from social platforms and perform data preprocessing operations to model the online social network as a heterogeneous graph. Assign category labels of social robots or real users to some user nodes based on manual annotation. Step 2, Encoding Multimodal Attributes for Feature Initialization: The preprocessed user metadata is divided into three categories: numerical attributes, boolean attributes, and text content. Initial feature vectors are generated through parametric feature transformation and processing by the pre-trained language model RoBERTa. And construct a directed social graph based on attention relationships. ,in, It is a set of nodes, where each node represents a user of a social media platform. To describe the edge set of user attention relationships, and thus based on and Train the CNMBot model to detect whether a user is a social bot based on their social relationships and attribute features; Step 3: Construct a community-aware view, a noise-aware view, and a motif-aware view, and obtain node representations under different views: based on the directed social graph. and initial feature vector From the perspectives of community discovery, relationship denoising, and behavioral collaboration, a community perception view, a noise perception view, and a motif perception view are constructed respectively, thereby obtaining the node representation under the community perception view. Node representation under the relationship denoising view Node representation under Motif-aware view To characterize the hierarchical community organizational structure among users, the reliability of interaction relationships, and the high-level collaborative behavior characteristics of user groups; Step 4: Adaptively fuse node features from different views for classification prediction, and use contrastive learning as an auxiliary task to optimize node representation: Employ a gated attention mechanism to calculate the gate signal for each view. By adaptively fusing node features from different views through element-wise multiplication, a fused comprehensive representation is obtained. ,Will The input is a multilayer perceptron for classification prediction, and the output is the prediction result of the corresponding node as a social robot or a human user. An auxiliary contrastive learning task with the noisy perception view as the anchor point is introduced, and the classification loss, contrastive loss and diffusion reconstruction loss are combined for joint training.

[0030] Optionally, step 1 includes modeling the online social network as a heterogeneous graph structure. The task of detecting social bots on social platforms is described as a binary classification problem of graph nodes, where It is a set of user nodes, each node representing a social media account, with a total of [number missing] nodes. ; To describe the edge set of interactions between users, It is a set of relationship types, which includes at least the user's follow and follow-up relationships; It is the metadata matrix of the nodes, where Representing the Metadata for each node, including the user's numerical statistical characteristics. Boolean properties and text content , Represents the set of node labels, where, Instructing users Is it a bot account? This indicates the account is a bot, 0 represents a human, and the ultimate goal is to learn a mapping function. This allows each node to be categorized based on social network relationships and user attributes. Mapped to its corresponding category label.

[0031] In an embodiment of the present invention, step 2 specifically includes: Step 21, for each user Considering the properties of the three modalities, namely numerical statistical characteristics Boolean properties and text content In this process, the text content is encoded using a pre-trained language model, RoBERTa, to capture semantic information, while numerical and categorical features are preprocessed using a learnable projection matrix. and Each feature is converted into a vector through linear mapping, and then concatenated with the three types of feature representations mentioned above before being input into a multilayer perceptron for nonlinear fusion processing to obtain the nodes. initial features , Formula 1; in, and It is a projection matrix. Represents the language model encoder, Indicates splicing operation Step 22, in the relationship graph construction section, a directed graph is constructed based on the follow and follow relationships between users. The constructed graph will be compared with the initial feature matrix Together, they serve as inputs for subsequent multi-view learning modules to construct the CNMBot model.

[0032] In an embodiment of the present invention, step 3 specifically includes: Step 31: Construct a community-aware view to capture the global hierarchical community structure, identify potential social robot clusters obscured by structural noise and dummy connections, and learn node representations. ; Step 32: Construct a noise-aware view by modeling edge reliability through a generative diffusion process to filter out unreliable and spoofed relationships, and learn node representations. ; Step 33, constructing the Motif-aware view, involves explicitly modeling the collaborative interaction patterns of the social robot by mining higher-order patterns, and learning node representations. .

[0033] In an embodiment of the present invention, step 31 specifically includes: Step 311, specify the tree height as By minimizing the social graph structural entropy To construct a hierarchical coding tree ; Step 312: Utilize a structural entropy pooling network to aggregate node features layer by layer along the encoding tree, capturing global hierarchical community patterns masked by structural noise, and obtaining node representations under the community-aware view. , Formula 2; in, This represents the hierarchical aggregation function of SEP-N. These are the parameters of the structural entropy pooling network. The initial node features are used; the mechanism gradually abstracts the features from the local neighborhood to the global community context, generating a high-order representation that is robust to structural perturbations.

[0034] In an embodiment of the present invention, step 32 specifically includes: Step 321: Using a denoised diffusion probability model, the observed social edges are treated as latent variables containing noise. The forward process is in Gaussian noise is gradually injected into the edge variables within each time step, converting the discrete edge indicators into continuous noise observations. The reverse process involves introducing a learnable denoiser. Embedded by edges and time step embedding To conditionally predict noise, minimize the difference between the predicted noise and the actual noise. Optimize the denoiser based on the mean square error between them ; Step 322, from the time step pure Gaussian noise distribution We begin by reconstructing edges using the learned denoiser and then applying sigmoid activation to calculate confidence scores. A weighted mask is generated using confidence scores. ; Step 323: Aggregate information on reliable social relationships through a relational graph convolutional network to obtain node representations under a noise-aware view. , Formula 3; in, Indicates by parameters The message passing function is driven by this diffusion denoising mechanism, which allows the encoder to focus more on aggregating information from reliable social relationships by evaluating the reliability of the edges.

[0035] In an embodiment of the present invention, step 33 specifically includes: Step 331 constructs three semantic adjacency matrices corresponding to the basic structure Motif. One of the first-order connections is a Motif Capturing star-shaped structure and fundamental connectivity, constructed using an undirected adjacency matrix: Two-way interactive Motif By identifying bidirectional edges to model mutual social trust, and constructing the Hadamard product: ; Ternary closure Motif Capturing community cohesion and coordinated interactions through the use of binary matrices Extracting triangular structures from user interactions: ; Step 332: Employ a multi-channel graph convolutional network to aggregate information within these semantic spaces, for each motif The node feature update is shown in Equation 4. Equation 4; Step 333: Apply an attention mechanism to aggregate the embeddings learned from the three channels to obtain the node representations under the final Motif-aware view. , Formula 5; in, This represents the Motif aggregation function, which combines Motif convolution with attention fusion, and is defined by parameters. The mechanism learns higher-order representations by modeling collaborative behaviors among users as motifs.

[0036] In an embodiment of the present invention, step 4 specifically includes: Step 41: A gating attention mechanism is designed to adaptively fuse node features from three views to obtain the fused node representation. , (Right now The (line) is the user Feature representation after fusing different views; Step 42, Node Representation The data is fed into a multilayer perceptron to predict the probability that each user is a social bot; for each node... , Represents the model for nodes The predicted probability of belonging to the category, Formula 6; The classifier is trained in a supervised manner using binary cross-entropy loss, and the classification loss is... Calculated using Equation 7 Formula 7; in, Represents the set of training nodes. For nodes The true label; Step 43 introduces self-supervised contrastive learning as an auxiliary task to align the semantics between different views and enhance the robustness of node representations; noise-aware views are then used. When specified as an anchor view, the goal of contrastive learning is to maximize the anchor view. With two complementary views and Mutual information between them; two alignment items and Calculated using Equations 8 and 9 Formula 8; Equation 9; Among them, losses Use InfoNCE loss; Step 44: Perform multi-objective joint optimization to calculate the final loss of CNMBot. It is a weighted combination of classification, comparison, and diffusion reconstruction losses. Formula 10; in, , , These are hyperparameters for balancing auxiliary objectives. For binary cross-entropy classification loss, and The InfoNCE loss represents the relationship between the representations learned by the community-aware view, the motif-aware view, and the noise-aware view, respectively. This represents the reconstruction mean square error loss for the diffusion model.

[0037] In an embodiment of the present invention, step 41 specifically includes: Step 411, for each view , where the view Represents a community-aware view, view Represents a noise-perceived view, view Representing the Motif-sensing view, the gating signal is calculated using Equation 11. , Formula 11; in, This represents the gating signal vector corresponding to view k, used to adaptively measure the contribution of that view to node identity determination. This represents the learnable weight matrix corresponding to view k. This represents the learnable bias vector corresponding to view k. For activation functions; Step 412: Element-wise multiply the features learned from each view with the gating signal to obtain the fused node representation. , Equation 12; in, This represents the Hadamard product.

[0038] The present invention will be described in detail below with reference to the embodiments and accompanying drawings.

[0039] Example 1: The social robot detection method based on multi-view graph representation learning according to embodiments of the present invention. Figure 1 A flowchart illustrating the social robot detection method based on multi-view graph representation learning according to this embodiment is shown. Figure 2 This diagram illustrates the overall framework of the CNMBot model constructed according to the multi-view graph representation learning-based social robot detection method of this embodiment. (See also...) Figure 1 and Figure 2 The methods include: Step 1: Extract user metadata and social relationship data from the social platform and perform data preprocessing operations to model the online social network as a heterogeneous graph structure. The task of detecting social bots on social platforms is described as a binary classification problem of graph nodes, where It is a set of user nodes, each node representing a social media account, with a total of [number missing] nodes. ; To describe the edge set of interactions between users, It is a set of relationship types, which includes at least the user's follow and follow-up relationships; It is the metadata matrix of the nodes, where, Representing the Metadata for each node, including the user's numerical statistical characteristics. (e.g., number of followers), Boolean attributes (such as authentication status) and text content (Tweets and descriptions) Based on manual annotation, some user nodes are assigned category labels as either social bots or real users. Represents the set of node labels, where, Instructing users Is it a bot account? This indicates the account is a bot, 0 represents a human, and the ultimate goal is to learn a mapping function. This allows each node to be categorized based on social network relationships and user attributes. Mapped to its corresponding category label.

[0040] Step 2, Feature Initialization. The preprocessed user metadata is divided into three categories: numerical attributes, boolean attributes, and text content. Initial feature vectors are generated using learnable feature projection transformation and the pre-trained language model RoBERTa. And construct a directed social graph based on attention relationships. ,in, It is a set of nodes, where each node represents a user. To describe the set of edges representing user attention relationships, and thus based on and Train the CNMBot model.

[0041] Specifically, the following steps are included: Step 21, for each user Considering the properties of the three modalities, namely numerical statistical characteristics Boolean properties and text content In this process, the text content is encoded using a pre-trained language model, RoBERTa, to capture semantic information, while numerical and categorical features are preprocessed using a learnable projection matrix. and Each feature is linearly mapped to a vector, and then concatenated with the three types of feature representations mentioned above before being input into a multilayer perceptron (MLP) for nonlinear fusion processing to obtain the node. initial features , Formula 1; in, and It is a projection matrix. Represents the language model encoder, Indicates a splicing operation; Step 22, in the relationship graph construction section, a directed graph is constructed based on the follow and follow relationships between users. The constructed graph will be compared with the initial feature matrix Together, they serve as inputs for subsequent multi-view learning modules to construct the CNMBot model.

[0042] Step 3, Multi-view Graph Representation Learning. Based on the directed social graph... and initial feature vector From the perspectives of community discovery, relationship denoising, and behavioral collaboration, a community perception view, a noise perception view, and a motif perception view are constructed respectively, thereby obtaining the node representation under the community perception view. Node representation under the relationship denoising view Node representation under Motif-aware view This is to characterize the hierarchical community organizational structure among users, the reliability of their interaction relationships, and the high-level collaborative behavior characteristics of user groups.

[0043] Specifically, the following steps are included: Step 31: Construct a community-aware view to capture the global hierarchical community structure, identify potential social robot clusters obscured by structural noise and dummy connections, and learn node representations. .

[0044] Step 311, specify the tree height as By minimizing the social graph structural entropy To construct a hierarchical coding tree .

[0045] like Figure 3 As shown, the height of the coding tree Taking an example, this paper illustrates the construction process of the coding tree in the community-aware view and its corresponding cluster allocation matrix. The construction process of the coding tree is based on the principle of minimizing structural entropy, revealing the inherent hierarchical community organization of the network by reducing the uncertainty of cross-level information flow in the social network. In this coding tree structure, leaf nodes directly correspond to user accounts on the social platform, while internal nodes represent hierarchical communities with different levels of abstraction. Nodes under the same branch of the coding tree are classified into the same community. This hierarchical structure can effectively capture the topological evolution from local sub-communities to global large groups. The cluster allocation matrix is ​​used to accurately encode the subordinate mapping relationship between user nodes and communities at each level, thereby clearly defining the community affiliation of each user at different structural scales.

[0046] Step 312: Utilize the Structural Entropy Pooling Network (SEP-N) to aggregate node features layer by layer along the encoding tree, capturing global hierarchical community patterns masked by structural noise, and obtaining node representations under the community-aware view. , Formula 2; in, This represents the hierarchical aggregation function of SEP-N. The parameters are those of the Structured Entropy Pooling Network (SEP-N). The initial node features are used. This mechanism progressively abstracts the features from the local neighborhood to the global community context, generating a higher-order representation that is robust to structural perturbations.

[0047] Step 32: Construct a noise-aware view by modeling edge reliability through a generative diffusion process to filter out unreliable and spoofed relationships, and learn node representations. .

[0048] Step 321: Using the Denoising Diffusion Probability Model (DDPM), the observed social edges are treated as latent variables containing noise. The forward process is in Gaussian noise is gradually injected into the edge variables within each time step, converting the discrete edge indicators into continuous noise observations. The reverse process involves introducing a learnable denoiser. Embedded by edges and time step embedding To conditionally predict noise, minimize the difference between the predicted noise and the actual noise. The denoiser is optimized based on the mean square error (MSE) between the two sides. .

[0049] Step 322, from the time step pure Gaussian noise distribution We begin by reconstructing edges using the learned denoiser and then applying sigmoid activation to calculate confidence scores. A weighted mask is generated using confidence scores. ; Step 323: Aggregate information on reliable social relationships through a Relationship Graph Convolutional Network (RGCN) to obtain node representations under a noise-aware view. , Formula 3; in, Indicates by parameters The message passing function is driven by this diffusion denoising mechanism. By evaluating the reliability of edges, the encoder can focus more on aggregating information from reliable social relationships.

[0050] Step 33, constructing the Motif-aware view, involves explicitly modeling the collaborative interaction patterns of the social robot by mining higher-order patterns, and learning node representations. .

[0051] Step 331 constructs three semantic adjacency matrices corresponding to the basic structure Motif. One of the first-order connections is a Motif Capturing star-shaped structure and fundamental connectivity, constructed using an undirected adjacency matrix: Two-way interactive Motif By identifying bidirectional edges to model mutual social trust, and constructing the Hadamard product: ; Ternary closure Motif Capturing community cohesion and coordinated interactions through the use of binary matrices Extracting triangular structures from user interactions: ; Step 332: Employ a multi-channel graph convolutional network to aggregate information within these semantic spaces, for each motif The node feature update is shown in Equation 4: Equation 4; in, Indicates the first In a convolutional layer, for a specific structural motif The learned node feature representation matrix; σ( () represents a non-linear activation function, used to enhance the model's feature extraction and non-linear fitting capabilities; This represents the semantic adjacency matrix constructed in step 331, which corresponds to specific higher-order structural patterns such as first-order attention, bidirectional attention, or ternary closure. Representation and semantic adjacency matrix The inverse of the corresponding degree matrix is ​​used to normalize the aggregated node features to ensure the stability of numerical calculations. Indicates the first For the phantom in the convolutional layer The node feature input, where at layer 0, The initial node feature matrix ; Indicates the first For the phantom in the convolutional layer The learnable weight matrix is ​​used to perform linear transformations and mappings on the structural information aggregated within the neighborhood.

[0052] Step 333: Apply an attention mechanism to aggregate the embeddings learned from the three channels to obtain the node representations under the final Motif-aware view. , Formula 5; in, This represents the Motif aggregation function, which combines Motif convolution with attention fusion, and is defined by parameters. The decision is made by learning higher-order representations through modeling collaborative behaviors among users as motifs.

[0053] Step 4, Detection and Optimization. A gating attention mechanism is used to calculate the gating signal for each view. By adaptively fusing node features from different views through element-wise multiplication, a fused comprehensive representation is obtained. ,Will The input is a multilayer perceptron (MLP) for classification prediction, and the output is the prediction result of whether the corresponding node is a social robot or a human user. An auxiliary contrastive learning task with the noisy perception view as the anchor point is introduced, and the classification loss, contrastive loss and diffusion reconstruction loss are jointly trained to improve the discriminativeness and stability of the classification decision.

[0054] Specifically, the following steps are included: Step 41: A gating attention mechanism is designed to adaptively fuse node features from three views to obtain the fused node representation. , (Right now The (line) is the user The feature representation after fusing different views.

[0055] Step 411, for each view , where the view Represents a community-aware view, view Represents a noise-perceived view, view Representing the Motif-sensing view, the gating signal is calculated using Equation 11. , Formula 11; in, This represents the gating signal vector corresponding to view k, used to adaptively measure the contribution of that view to node identity determination. This represents the learnable weight matrix corresponding to view k, which is used to perform linear transformations on the view features. This represents the learnable bias vector corresponding to view k, used to adjust the offset of the gating signal; This is the activation function, used to introduce nonlinear mappings to enhance the gating mechanism.

[0056] Step 412: Element-wise multiply the features learned from each view with the gating signal to obtain the fused node representation. , Equation 12; in, This represents the Hadamard product.

[0057] Step 42, Node Representation The data is fed into a multilayer perceptron (MLP) to predict the probability that each user is a social bot. For each node... , Represents the model for nodes The predicted probability of belonging to the category, Formula 6; The classifier is trained in a supervised manner using binary cross-entropy loss, and the classification loss is... Calculated using Equation 7 Formula 7; in, Represents the set of training nodes. For nodes The true label; Step 43 introduces self-supervised contrastive learning as an auxiliary task to align the semantics between different views and enhance the robustness of node representations. This involves handling noise-aware views. When specified as an anchor view, the goal of contrastive learning is to maximize the anchor view. With two complementary views and Mutual information between the two alignment items. and Calculated using Equations 8 and 9 Formula 8; Equation 9; Among them, losses Use InfoNCE loss; Step 44: Perform multi-objective joint optimization to calculate the final loss of CNMBot. It is a weighted combination of classification, comparison, and diffusion reconstruction losses. Formula 10; in, , , These are hyperparameters for balancing auxiliary objectives. For binary cross-entropy classification loss, and The InfoNCE loss represents the relationship between the representations learned by the community-aware view, the motif-aware view, and the noise-aware view, respectively. This represents the reconstruction mean square error loss for the diffusion model.

[0058] The social robot detection method based on multi-view graph representation learning provided by this invention was applied to publicly available social robot detection datasets to verify the performance of the method. The selected datasets were TwiBot-20 and MGTAB. The statistical results of the datasets are shown in Table 1 below.

[0059] Table 1 Attributes of the Experimental Dataset

[0060] The social robot detection method based on multi-view graph representation learning provided by this invention is compared with 10 existing methods. The 10 selected benchmark methods are shown below: (1) GCN: A classic graph neural network model that updates node representations by aggregating information from neighboring nodes.

[0061] (2) GAT: Introduces an attention mechanism during feature aggregation to adaptively assign weights to neighboring nodes.

[0062] (3) GBT: It uses BarlowTwins loss as a self-supervised signal, without the need for negative sampling operation and asymmetric architecture of traditional contrastive learning.

[0063] (4) FAGCN: Adaptively integrates different signals during message passing through a self-gating mechanism.

[0064] (5) HGT: Using relation-aware multi-head attention and relative temporal encoding to capture the heterogeneity and dynamic dependencies of social graphs.

[0065] (6) RGT: Modeling heterogeneous information networks using relational graph Transformer and attention mechanism to handle the heterogeneity of relations and influence strength in online networks.

[0066] (7) BotRGCN: Utilizes relational graph convolutional networks to model and learn various relationships between users.

[0067] (8) SEBot: Introduces structural entropy to learn the potential hierarchical community structure and captures more comprehensive semantic information through subgraph-level and node-level comparative learning.

[0068] (9) OS GNN: It solves the class imbalance problem by generating synthetic nodes for the minority class, thereby improving minority class performance and overall accuracy in imbalanced graphs.

[0069] (10) BotSTIP: Considering the temporal dynamics of social networks, it constructs a hypergraph by performing community detection in each time interval and uses the Transformer module to extract temporal information, thereby effectively modeling dynamic robot behavior.

[0070] In the experiment, the hidden feature dimension was set to 32. The depth of the encoding tree in the community-aware view was set to 6 to balance accuracy and computational cost. The phantom-aware view used two graph convolutional layers. In contrastive learning, the projector head converted the feature dimension to 16, and the weight coefficients of both terms in the contrastive loss were set to 0.05. In the noise-aware view, the number of diffusion steps was fixed at 5. Other diffusion-related hyperparameters need to be adjusted for specific datasets: the diffusion loss coefficient is set in TwiBot. Set to 20 Set on MGTAB Similarly, negative sampling rates in TwiBot The learning rate was set to 0.5 on 20 and 0.7 on MGTAB. Model parameters were optimized using the AdamW optimizer with an initial learning rate of 0.01. Accuracy, Precision, and F1 score were also used. The scores and recall rates are used as metrics. Higher values ​​for all metrics indicate better detection performance.

[0071] Tables 2 and 3 show the comparison of CNMBot with benchmark methods in TwiBot. A comparison of overall performance on 20 and MGTAB shows the best results highlighted in bold, while the second-best results are highlighted with an underline. "-" indicates that the corresponding method is not scalable or cannot be applied due to missing raw data. Based on Tables 2 and 3, it can be observed that CNMBot performs better on TwiBot. It consistently achieves best or near-best performance across all metrics on the 20 and MGTAB datasets. (In TwiBot) On TwiBot 20, CNMBot achieved the highest accuracy and precision, while also achieving competitive F1 scores and recall. On MGTAB, CNMBot outperformed all baselines in accuracy and F1 score, demonstrating its strong robustness across datasets of varying sizes and annotation quality. Furthermore, several baseline methods showed strong performance on specific metrics or datasets. For example, RGT performed well on TwiBot. BotSTIP performed well on TwiBot with an accuracy of 20 and an F1 score. High recall rate was achieved on 20, OS GNNs demonstrate competitive accuracy on MGTAB. However, their performance varies significantly across different datasets and evaluation metrics. This inconsistency suggests that models relying on a single structural assumption or a limited representational view are insufficient to capture the full range of behavioral patterns of social bots, thus limiting their robustness to the structural heterogeneity and relational spoofing prevalent in real-world social networks. The above analysis demonstrates that the method of this invention can effectively improve the performance of social bot detection.

[0072] Table 2 TwiBot Performance comparison of different methods on 20 datasets

[0073] Table 3 Performance comparison of different methods on the MGTAB dataset

[0074] Ablation experiments were conducted by removing specific modules. Five variants were defined as follows: w / oMotif removes the motif-aware view; w / oDenoising removes the noise-aware view; w / oSEP removes the community-aware view; w / oFusion replaces gated fusion with simple stitching; and w / oContrastive removes the auxiliary contrastive learning task. Experimental results are shown in Tables 4 and 5.

[0075] Table 4 TwiBot Performance comparison of different variants on 20 datasets

[0076] Table 5 Performance comparison of different variants on the MGTAB dataset

[0077] As can be observed from Tables 4 and 5, the complete CNMBot model consistently achieves the best performance, while the performance of the model decreases after removing some components, indicating that all components contribute to the final detection accuracy. In particular, removing the diffusion denoising module leads to a significant performance drop, for example, in TwiBot. Accuracy decreased by 1.01% on 20 and by 1.58% on MGTAB, highlighting the importance of suppressing noisy edges introduced by social robots. Removing the phantom-aware view resulted in a significant drop in F1 score and recall, indicating that modeling higher-order structural patterns is beneficial for capturing collaborative robot behavior. The community-aware view and contrastive learning module also contribute to the stability and robustness of the framework. Removing the community-aware view prevents the model from perceiving hierarchical community structures, while removing contrastive learning weakens the discriminative power of the learned representations. Furthermore, replacing the gated fusion mechanism with simple concatenation leads to a continuous performance decline, particularly in accuracy and F1 score. This demonstrates that adaptive gated aggregation is more effective than simple fusion in integrating complementary information from multiple views.

[0078] Figure 4 and Figure 5 These are experimental comparison figures showing the impact of different negative sampling rates and diffusion steps on the performance metrics of the proposed method on the TwiBot-20 and MGTAB datasets. Figure 4 As shown, negative sampling rate The change from 0.1 to 0.9 can be observed in TwiBot. On dataset 20, when At that time, the proposed model reached its peak values ​​on all metrics. Further increasing the negative sampling rate resulted in a slight performance degradation, indicating that excessive noise might mask meaningful structural signals. For MGTAB, the model... It achieves peak accuracy and demonstrates competitive performance on other metrics. These observations suggest that while a balanced noise ratio is generally effective, denser plots like MGTAB may benefit from a slightly higher noise ratio to enhance the robustness of the denoiser. Figure 5 As shown, the method performance varies with the number of diffusion steps. Increase and improve, and in At TwiBot On 20 and MGTAB, the results show optimal or near-optimal values ​​for most metrics. This indicates that a sufficient number of diffusion steps is necessary to separate structural noise from the true connection. However, when T becomes too large (e.g., ...), ... When the model's performance deteriorates on both datasets, it may be due to the over-smoothing effect disrupting the true connection structure.

[0079] Figure 6 This is an experimental comparison chart showing the impact of different coding tree depths on various performance metrics of the method of this invention on the TwiBot-20 and MGTAB datasets. (See figure below.) Figure 6 As shown, when the depth of the coding tree At that time, the model was in TwiBot Both 20 and MGTAB consistently achieve optimal or near-optimal performance across all metrics. When the tree is shallow ( When the depth is increased beyond this point, the model cannot fully capture multi-level semantics, resulting in suboptimal representation. Conversely, increasing the depth beyond this point... This approach will not lead to further performance improvements and may even introduce redundant partitions and additional noise, while generating unnecessary computational overhead. The above results demonstrate that an appropriate tree depth provides the optimal balance between expressive hierarchical modeling and model efficiency.

[0080] Figure 7 Different contrast loss coefficients and A heatmap showing the change in the detection accuracy of the method of the present invention under combined conditions, and Figure 8 Different diffusion loss coefficients Experimental comparison of the impact of the method of this invention on various performance metrics on the TwiBot-20 and MGTAB datasets.

[0081] like Figure 7 As shown, in TwiBot On 20 and MGTAB, when the loss coefficients of both comparison terms are set to... and When the comparison weights are too small (e.g., ...), the model can always achieve optimal or near-optimal performance. When this happens, the auxiliary target cannot effectively force cross-viewing. Figure 1 Consistency leads to suboptimal representation. Conversely, if excessive weights are assigned (e.g., ... , Auxiliary tasks often dominate the optimization process, obscuring the gradient direction of the main classification loss and thus hindering model convergence. The above analysis illustrates that appropriate contrastive supervision plays a crucial role in enhancing the discriminability of representations without compromising training stability. For example... Figure 8 As shown, in TwiBot On 20, the model is It achieves optimal or near-optimal performance across all metrics, while larger MGTAB datasets tend to have smaller weights. The above changes indicate that the impact of reconstructing the target should be adjusted according to the characteristics of the dataset to prevent interference with the main detection task. Furthermore, a smaller diffusion loss coefficient can be observed. This will weaken the effect of diffusion denoising, limiting the model's ability to suppress noise or masquerade edges. Conversely, excessively large... This high value forces the model to prioritize structural reconstruction over the core robot detection task, resulting in a significant drop in accuracy and F1 score. Overall, these results illustrate the importance of carefully balancing the denoising objective with the classification loss for maintaining the performance of the method described in this invention.

[0082] Figure 9 This is a visualization comparing the clustering results of node representations learned by different methods on the TwiBot-20 dataset after processing with t-SNE. It shows a comparison between CNMBot and several representative baselines, where red squares represent social robot nodes and blue circles represent human nodes. Figure 9 As shown in (a) and (b), the node embeddings generated by GCN and GAT exhibit significant overlap between humans (blue) and robots (red), indicating that such homogeneous GNNs struggle to capture the behavioral patterns of social robots in complex social networks. Heterogeneous GNN models, such as HGT and RGT (…),… Figure 9 (c) and (d) improved the feature separation between bot accounts and human accounts to some extent; however, the clustering is still fragmented, and the category boundaries remain blurred. BotRGCN, SEBot, and OS GNN ( Figure 9 (e) Methods such as (g) further enhance separation by incorporating structural or semantic elements, but significant inter-class mixing and dispersed distribution still exist, indicating limited robustness to structural noise and masquerading interactions. In contrast, CNMBot ( Figure 9 (h) exhibits clear and compact clustering with well-defined decision boundaries between humans and robots. This result demonstrates the effectiveness of the multi-view design in the method of this invention, capturing the global hierarchical structure through community perception, suppressing noisy relationships through noise perception, and utilizing high-order cooperative patterns through phantom perception modeling, thereby obtaining a comprehensive node representation to distinguish between social robots and human users.

[0083] Although the present invention has been described in detail above with general descriptions and specific embodiments, modifications or improvements can be made to it, which will be obvious to those skilled in the art. Therefore, all such modifications or improvements made without departing from the spirit of the present invention fall within the scope of protection claimed by the present invention.

Claims

1. A social robot detection method based on multi-view graph representation learning, characterized in that, Includes the following steps: Step 1: Obtain user data from social platforms and construct a heterogeneous graph: Extract user metadata and social relationship data from social platforms and perform data preprocessing operations to model the online social network as a heterogeneous graph. Assign category labels of social robots or real users to some user nodes based on manual annotation. Step 2, Encoding Multimodal Attributes for Feature Initialization: The preprocessed user metadata is divided into three categories: numerical attributes, boolean attributes, and text content. Initial feature vectors are generated through parametric feature transformation and processing by the pre-trained language model RoBERTa. And construct a directed social graph based on attention relationships. ,in, It is a set of nodes, where each node represents a user of a social media platform. To describe the edge set of user attention relationships, and thus based on and Train the CNMBot model to detect whether a user is a social bot based on their social relationships and attribute features; Step 3: Construct a community-aware view, a noise-aware view, and a motif-aware view, and obtain node representations under different views: based on the directed social graph. and initial feature vector From the perspectives of community discovery, relationship denoising, and behavioral collaboration, a community perception view, a noise perception view, and a motif perception view are constructed respectively, thereby obtaining the node representation under the community perception view. Node representation under the relationship denoising view Node representation under Motif-aware view To characterize the hierarchical community organizational structure among users, the reliability of interaction relationships, and the high-level collaborative behavior characteristics of user groups; Step 4: Adaptively fuse node features from different views for classification prediction, and use contrastive learning as an auxiliary task to optimize node representation: Employ a gated attention mechanism to calculate the gate signal for each view. By adaptively fusing node features from different views through element-wise multiplication, a fused comprehensive representation is obtained. ,Will The input is a multilayer perceptron for classification prediction, and the output is the prediction result of the corresponding node as a social robot or a human user. An auxiliary contrastive learning task with the noisy perception view as the anchor point is introduced, and the classification loss, contrastive loss and diffusion reconstruction loss are combined for joint training.

2. The social robot detection method based on multi-view graph representation learning as described in claim 1, characterized in that, Step 1 includes modeling the online social network as a heterogeneous graph structure. The task of detecting social bots on social platforms is described as a binary classification problem of graph nodes, where It is a set of user nodes, each node representing a social media account, with a total of [number missing] nodes. ; To describe the edge set of interactions between users, It is a set of relationship types, which includes at least the user's follow and follow-up relationships; It is the metadata matrix of the nodes, where Representing the Metadata for each node, including the user's numerical statistical characteristics. Boolean properties and text content , Represents the set of node labels, where, Instructing users Is it a bot account? This indicates the account is a bot, 0 represents a human, and the ultimate goal is to learn a mapping function. This allows each node to be categorized based on social network relationships and user attributes. Mapped to its corresponding category label.

3. The social robot detection method based on multi-view graph representation learning as described in claim 1 or 2, characterized in that, Step 2 specifically includes: Step 21, for each user Considering the properties of the three modalities, namely numerical statistical characteristics Boolean properties and text content In this process, the text content is encoded using a pre-trained language model, RoBERTa, to capture semantic information, while numerical and categorical features are preprocessed using a learnable projection matrix. and Each feature is converted into a vector through linear mapping, and then concatenated with the three types of feature representations mentioned above before being input into a multilayer perceptron for nonlinear fusion processing to obtain the nodes. initial features , Formula 1; in, and It is a projection matrix. Represents the language model encoder, Indicates a splicing operation; Step 22, in the relationship graph construction section, a directed graph is constructed based on the follow and follow relationships between users. The constructed graph will be compared with the initial feature matrix Together, they serve as inputs for subsequent multi-view learning modules to construct the CNMBot model.

4. The social robot detection method based on multi-view graph representation learning as described in claim 1 or 2, characterized in that, Step 3 specifically includes: Step 31: Construct a community-aware view to capture the global hierarchical community structure, identify potential social robot clusters obscured by structural noise and dummy connections, and learn node representations. ; Step 32: Construct a noise-aware view by modeling edge reliability through a generative diffusion process to filter out unreliable and spoofed relationships, and learn node representations. ; Step 33, constructing the Motif-aware view, involves explicitly modeling the collaborative interaction patterns of the social robot by mining higher-order patterns, and learning node representations. .

5. The social robot detection method based on multi-view graph representation learning as described in claim 4, characterized in that, Step 31 specifically includes: Step 311, specify the tree height as By minimizing the social graph structural entropy To construct a hierarchical coding tree ; Step 312: Utilize a structural entropy pooling network to aggregate node features layer by layer along the encoding tree, capturing global hierarchical community patterns masked by structural noise, and obtaining node representations under the community-aware view. , Formula 2; in, This represents the hierarchical aggregation function of SEP-N. These are the parameters of the structural entropy pooling network. The initial node features are used; the mechanism gradually abstracts the features from the local neighborhood to the global community context, generating a high-order representation that is robust to structural perturbations.

6. The social robot detection method based on multi-view graph representation learning as described in claim 4, characterized in that, Step 32 specifically includes: Step 321: Using a denoised diffusion probability model, the observed social edges are treated as latent variables containing noise. The forward process is in Gaussian noise is gradually injected into the edge variables within each time step, converting the discrete edge indicators into continuous noise observations. The reverse process involves introducing a learnable denoiser. Embedded by edges and time step embedding To conditionally predict noise, minimize the difference between the predicted noise and the actual noise. Optimize the denoiser based on the mean square error between them ; Step 322, from the time step pure Gaussian noise distribution We begin by reconstructing edges using the learned denoiser and then applying sigmoid activation to calculate confidence scores. A weighted mask is generated using confidence scores. ; Step 323: Aggregate information on reliable social relationships through a relational graph convolutional network to obtain node representations under a noise-aware view. , Formula 3; in, Indicates by parameters The message passing function is driven by this diffusion denoising mechanism, which allows the encoder to focus more on aggregating information from reliable social relationships by evaluating the reliability of the edges.

7. The social robot detection method based on multi-view graph representation learning as described in claim 4, characterized in that, Step 33 specifically includes: Step 331 constructs three semantic adjacency matrices corresponding to the basic structure Motif. One of the first-order connection motifs Capturing star-shaped structure and fundamental connectivity, constructed using an undirected adjacency matrix: Two-way interactive Motif By identifying bidirectional edges to model mutual social trust, and constructing the Hadamard product: ; Ternary closure Motif Capturing community cohesion and coordinated interactions through the use of binary matrices Extracting triangular structures from user interactions: ; Step 332: Employ a multi-channel graph convolutional network to aggregate information within these semantic spaces, for each motif The node feature update is shown in Equation 4. Equation 4; Step 333: Apply an attention mechanism to aggregate the embeddings learned from the three channels to obtain the node representations under the final Motif-aware view. , Formula 5; in, This represents the Motif aggregation function, which combines Motif convolution with attention fusion, and is defined by parameters. The mechanism learns higher-order representations by modeling collaborative behaviors among users as motifs.

8. The social robot detection method based on multi-view graph representation learning as described in claim 1 or 2, characterized in that, Step 4 specifically includes: Step 41: A gating attention mechanism is designed to adaptively fuse node features from three views to obtain the fused node representation. , (Right now The (line) is the user Feature representation after fusing different views; Step 42, Node Representation The data is fed into a multilayer perceptron to predict the probability that each user is a social bot; for each node... , Represents the model for nodes The predicted probability of belonging to the category, Formula 6; The classifier is trained in a supervised manner using binary cross-entropy loss, and the classification loss is... Calculated using Equation 7 Formula 7; in, Represents the set of training nodes. For nodes The true label; Step 43 introduces self-supervised contrastive learning as an auxiliary task to align the semantics between different views and enhance the robustness of node representations; noise-aware views are then used. When specified as an anchor view, the goal of contrastive learning is to maximize the anchor view. With two complementary views and Mutual information between them; two alignment items and Calculated using Equations 8 and 9 Formula 8; Equation 9; Among them, losses Use InfoNCE loss; Step 44: Perform multi-objective joint optimization to calculate the final loss of CNMBot. It is a weighted combination of classification, comparison, and diffusion reconstruction losses. Formula 10; in, , , These are hyperparameters for balancing auxiliary objectives. For binary cross-entropy classification loss, and The InfoNCE loss represents the relationship between the representations learned by the community-aware view, the motif-aware view, and the noise-aware view, respectively. This represents the reconstruction mean square error loss for the diffusion model.

9. The social robot detection method based on multi-view graph representation learning as described in claim 8, characterized in that, Step 41 specifically includes: Step 411, for each view , where the view Represents a community-aware view, view Represents a noise-perceived view, view Representing the Motif-sensing view, the gating signal is calculated using Equation 11. , Formula 11; in, This represents the gating signal vector corresponding to view k, used to adaptively measure the contribution of that view to node identity determination. This represents the learnable weight matrix corresponding to view k. This represents the learnable bias vector corresponding to view k. For activation functions; Step 412: Element-wise multiply the features learned from each view with the gating signal to obtain the fused node representation. , Equation 12; in, This represents the Hadamard product.

Images