Social robot detection system and method for false information in power market

Abstract

The embodiment of the present invention discloses a social robot detection system and method for false information in the power market. The system includes a construction unit and a detection unit. The construction unit constructs a HyperBot model of a social robot for publishing false fault information in the power market. The detection unit determines whether the user is a social robot or a real user through the HyperBot model. The HyperBot model includes a hypergraph construction module, a multimodal user network module, a hypergraph expert network module, a relationship fusion network module, and a feature fusion layer. The present invention not only helps to maintain the healthy ecology of social media platforms, but also provides important technical support for the field of information security, and especially enhances the monitoring capability of social robots' false fault reporting behavior in the power system, effectively protecting the stability of the power market.

Description

Technical Field

[0001] The present invention relates to the technical field of false information detection, and in particular to a social robot detection system and method for false information in the power market. Background Art

[0002] With the rapid development of the Internet, social media platforms have become an integral part of people's daily lives. These platforms provide users with a space for sharing and communication, but they have also become a hotbed for malicious actors. Among them, the proliferation of social bots has become a problem that needs to be solved urgently. Social bots simulate human user behavior in an automated way and are able to spread information on a large scale on social media. In recent years, criminals have begun to use social bots to falsely report faults in the power market, disrupt power supply, and further affect the stability of the power market. Therefore, effectively detecting and identifying these social bots, especially the spread of false information involving power systems, has become an important topic in the current information security field.

[0003] Existing social robot detection methods are mainly divided into two categories: those based on user behavior characteristics and those based on social network structure. Methods based on user behavior characteristics usually use features such as user tweet content, posting frequency, and interactive behavior to classify through machine learning or deep learning models. However, with the continuous advancement of social robot technology, these features are becoming increasingly difficult to distinguish. Advanced social robots are able to evade detection by mimicking the behavior patterns of human users and generating high-quality content. This makes methods that rely solely on behavioral features limited in effectiveness when facing complex and advanced social robots.

[0004] Methods based on social network structure use the relationship network between users and identify social robots by analyzing network structure features such as users' attention relationships and interaction frequency. The advantage of this type of method is that although social robots can disguise their behavior as human users, their relationship structure in social networks is often difficult to completely imitate real users. However, traditional network structure analysis methods are mostly based on simple graph structures and can only capture direct relationships between users, ignoring the complex high-order relationships in social networks.

[0005] In order to better capture the complex relationships in social networks, and pay special attention to social robot activities that may affect the power system, the hypergraph model has been introduced into the field of social robot detection. Hypergraphs are different from traditional simple graphs. They can represent complex interactive relationships between multiple users and are suitable for describing various types of relationships in social networks. For example, in a social network, multiple users follow a certain user, or one user follows multiple other users. These relationships can be modeled through hypergraphs. Through hypergraphs, the relationship structure in social networks can be more comprehensively represented, improving the accuracy of detection.

[0006] With the widespread use of social media platforms, the proliferation of social bots has become a growing problem. Social bots are now being used in power systems to post false fault reports to disrupt the power market. In order to effectively identify and detect these bots, researchers have proposed a variety of methods, which are mainly divided into three categories: feature-based, text-based, and graph-based. Although these methods have been successful to a certain extent, they still have some shortcomings in the application scenarios of power systems.

[0007] Feature-based methods:

[0008] Feature-based methods mainly use users’ metadata and behavioral features to detect social robots. For example, SGBot uses a random forest classifier to perform feature engineering on users’ metadata and derived features. The advantages of this type of method are that it is relatively simple to implement and computationally efficient. However, with the continuous advancement of social robot technology, many advanced robots are able to simulate the behavioral characteristics of real users, making methods that rely only on metadata and behavioral features have limited effectiveness when facing advanced social robots, especially in the false fault reporting scenario in power systems.

[0009] Text-based methods:

[0010] Text-based methods identify social bots by analyzing the content posted by users. For example, Wei et al. used a recurrent neural network (RNN) to capture features in tweets and transformed social bot detection into a text classification task. Such methods are able to capture complex patterns in text and thus identify automatically generated content. However, the text content generated by social bots is increasingly difficult to distinguish from human users, especially with the rapid development of natural language generation technology. Therefore, the recognition accuracy of text-based methods may decline when faced with social bots that generate high-quality text.

[0011] Graph-based methods:

[0012] Graph-based methods use the relationship structure between users in social networks for detection. For example, BotRGCN constructs a heterogeneous graph to represent the user network and uses a relational graph convolutional network (RGCN) for representation learning and robot detection. This type of method can capture the complex relationships between users by leveraging the structural information in social networks. However, traditional graph models can often only represent simple edge relationships and have difficulty capturing high-order complex relationships. In addition, the computational cost of building and processing large-scale graph structures is high, which may affect the real-time and scalability of the model, especially in power system fault detection that requires rapid response.

[0013] Although existing social robot detection methods have achieved some success in the field of social media, there are still several major problems in their application in power systems:

[0014] 1. Difficult to distinguish simulated behaviors: Advanced social robots are able to imitate the behavioral characteristics of human users and generate high-quality text content, making it difficult for feature-based and text-based methods to distinguish these robots from real users, especially in fault reporting scenarios in power systems.

[0015] 2. Ignoring high-order relationships: Traditional graph-based methods can usually only capture simple edge relationships between users and have difficulty representing and utilizing high-order complex relationships in social networks. This performs poorly when faced with advanced social robots with complex interaction patterns.

[0016] 3. High computational cost: Building and processing large-scale graph structures requires a lot of computing resources, which affects the real-time and scalability of the model, which is an important limitation for power systems that require fast and accurate responses.

[0017] 4. Insufficient fusion of multimodal information: User information on social media is usually multimodal, including text, pictures, videos, metadata, etc. Most existing methods focus on a certain type of information and it is difficult to fully integrate information from multiple modalities, which affects the detection effect, especially when analyzing technical fault reports. Summary of the invention

[0018] The technical problem to be solved by the embodiments of the present invention is to provide a social robot detection system and method for false information in the power market, so as to enhance the monitoring capability of the social robot's false fault reporting behavior in the power system and effectively protect the stability of the power market.

[0019] In order to solve the above technical problems, an embodiment of the present invention proposes a social robot detection system for false information in the power market, including a construction unit and a detection unit. The construction unit constructs a HyperBot model of a social robot for publishing false fault information in the power market. The detection unit inputs user information that publishes fault information in the power market into the HyperBot model, and determines whether the user is a social robot or a real user through the HyperBot model.

[0020] The HyperBot model includes a hypergraph construction module, a multimodal user network module, a hypergraph expert network module, a relationship fusion network module, and a feature fusion layer.

[0021] The hypergraph construction module constructs a hypergraph model, uses the hypergraph model to build a social network, and captures the friendship and attention relationships of users in the social network;

[0022] The multimodal user network module extracts and fuses features from the user's text content and metadata to obtain a multimodal representation of each user;

[0023] The hypergraph expert network module learns user relationships in simple graphs, hypergraphs, and composite graphs of simple graphs and hypergraphs;

[0024] The relationship fusion network module uses a multi-head self-attention mechanism to fuse different types of relationship features for user classification;

[0025] The feature fusion layer simultaneously learns the friend relationships and follow relationships associated with the user to classify the user as a social robot or a real user.

[0026] Accordingly, an embodiment of the present invention further provides a social robot detection method for false information in the power market, including a construction step and a detection step.

[0027] Construction steps: Construct a HyperBot model of a social robot used to publish false fault information in the power market; the HyperBot model includes a hypergraph construction module, a multimodal user network module, a hypergraph expert network module, a relationship fusion network module, and a feature fusion layer, among which,

[0028] The hypergraph construction module constructs a hypergraph model, uses the hypergraph model to build a social network, and captures the friendship and attention relationships of users in the social network;

[0029] The multimodal user network module extracts and fuses features from the user's text content and metadata to obtain a multimodal representation of each user;

[0030] The hypergraph expert network module learns user relationships in simple graphs, hypergraphs, and composite graphs of simple graphs and hypergraphs;

[0031] The relationship fusion network module uses a multi-head self-attention mechanism to fuse different types of relationship features for user classification;

[0032] The feature fusion layer simultaneously learns the friend relationships and follow relationships related to the user, and classifies the user as a social robot or a real user;

[0033] Detection steps: Input the user information of the user who publishes the fault information in the power market into the HyperBot model, and use the HyperBot model to determine whether the user is a social robot or a real user.

[0034] The beneficial effects of the present invention are as follows: by introducing a hypergraph model and multimodal feature representation, the present invention can more comprehensively capture complex relationships and user information in social networks, and provide a new solution for effectively detecting and identifying advanced social robots. The present invention not only helps maintain a healthy ecology of social media platforms, but also provides important technical support for the field of information security, and especially enhances the monitoring capability of social robots' false fault reporting behavior in power systems, effectively protecting the stability of the power market. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Schematic diagram of the structure of the HyperBot model of the embodiment of the present invention. DETAILED DESCRIPTION

[0036] It should be noted that, in the absence of conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present invention is further described in detail below in conjunction with the drawings and specific embodiments.

[0037] In the embodiments of the present invention, if there are directional indications (such as up, down, left, right, front, back, etc.), they are only used to explain the relative position relationship, movement status, etc. between the components under a certain specific posture (as shown in the accompanying drawings). If the specific posture changes, the directional indication will also change accordingly.

[0038] In addition, in the present invention, the descriptions of "first", "second", etc. are only used for descriptive purposes and cannot be understood as indicating or implying their relative importance or implicitly indicating the number of the indicated technical features. Therefore, the features defined as "first" or "second" may explicitly or implicitly include at least one of the features.

[0039] The social robot detection system for false information in the electricity market according to an embodiment of the present invention includes a construction unit and a detection unit.

[0040] The construction unit constructs and trains a HyperBot model of a social robot for publishing false fault information in the power market. The detection unit inputs user information that publishes fault information in the power market into the HyperBot model, and determines whether the user is a social robot or a real user through the HyperBot model.

[0041] Please refer to Figure 1 ,The HyperBot model includes a hypergraph building module, a multimodal user network module, a hypergraph expert network module, a relation fusion network module, and a feature fusion layer.

[0042] The hypergraph construction module constructs a hypergraph model, uses the hypergraph model to build a social network, and captures the friendship and attention relationships of users in the social network.

[0043] Hypergraph building blocks include:

[0044] The friend relationship hypergraph and the following relationship hypergraph represent the following and followed relationships of users respectively;

[0045] Friendship Super Chart , where each hyperedge represents a user and the accounts he follows;

[0046] Attention Relationship Hypergraph , where each hyperedge represents an account that multiple users follow together;

[0047] Supermap The correlation matrix It means that each element of the matrix Defined as:

[0048] ;

[0049] Among them, u represents the user in the user set, e represents the relationship edge connecting the users, Represents the elements in the H matrix.

[0050] The multimodal user network module extracts and fuses features from the user's text content and metadata to obtain a multimodal representation of each user. The multimodal user network module includes a description encoder, a tweet encoder, a category encoder, and a value encoder.

[0051] The description encoder is used to encode user descriptions, using a pre-trained language model (such as RoBERTa). The description encoding is represented as:

[0052] ;

[0053] in represents the pre-trained language model, represents the encoded features of the user description, express The embedding dimension, surface dimensional real space. Then, the present invention extracts the feature vector described by the user:

[0054] ;

[0055] Where W A and B A is a learnable parameter, is the user embedding dimension, express dimensional real space, Represents the activation function.

[0056] The tweet encoder is used to encode user tweets. Like user descriptions, the present invention uses the RoBERTa model pre-trained on the Twitter dataset to encode user tweets:

[0057] ;

[0058] in represents the encoded user tweet features, Corresponding to the embedding dimension of LM. Then, referring to the method describing the encoder, we get the user tweet features .

[0059] Category Encoder,For user profiles obtained through Twitter API, the present invention considers 3 simple and practical classification features (protected, verified, default profile picture). We use one-hot encoding for each classification feature and concatenate them to obtain a joint classification feature. By attaching a multi-layer perceptron (MLP) to learn the feature, the present invention obtains the category feature associated with the user .

[0060] Numeric encoder,Similar to the category features, this paper selects 5 numerical features (followers, following, status, active days, and username length) from the Twitter API. After z-score normalization and MLP, the numerical features of the corresponding user can be obtained. .

[0061] The present invention encodes and connects the descriptive features, tweet features, category features and numerical features to generate overall user features. , the present invention obtains the joint user characteristics:

[0062] ;

[0063] The present invention is used represents the joint feature set of all users, where .

[0064] The hypergraph expert network module learns user relationships in simple graphs, hypergraphs, and composite graphs of simple graphs and hypergraphs. The hypergraph expert network module contains multiple layers of hypergraph convolution and hybrid expert structures.

[0065] Multi-layer hypergraph convolution,After constructing the hypergraph and obtaining user features, it is essential to combine the two for modeling.,Inspired by previous work, this paper adopts Layer hypergraph convolution is used to aggregate and learn the feature information of the hypergraph structure. Indicates The learnable filter matrix of the layer hypergraph. The hypergraph convolution layer can be expressed as:

[0066] ;

[0067] in It is The signal of the layer hypergraph, , H is the hypergraph incidence matrix is ​​the hypergraph incidence matrix, represents the transposed matrix of the hypergraph incidence matrix, express The corresponding inverse matrix is, express The corresponding negative half-power matrix, Represents the weight of the hyperedge, which is preferably set to 1 in the present invention. and Represent the diagonal matrices used for normalization of vertex degrees and edge degrees, respectively. represents a non-linear activation function.

[0068] Mixture-of-Experts Structure,Although the convolutional layer can propagate high-order information through hyperedges, it does not consider the difference between simple and complex relations.,Inspired by the efficient training of Mixture-of-Experts (MoE) in large language models,,this paper introduces the MoE structure to learn neighborhood relations in social networks.

[0069] The present invention uses different expert models to learn simple and complex relationships. Specifically, for each type of relationship, the present invention constructs three hypergraph convolutional expert models to learn from simple graphs, hypergraphs and their combination. In order to understand the interaction and influence between individual and collective behaviors of users, the present invention integrates simple graphs with hypergraphs to form a composite graph. The combined graph of friend relationships and follow relationships is defined as:

[0070]

[0071] For a friendship relationship f, Representation and and have the same set of nodes The hypergraph of , including all simple edges and hyperedges in these graphs, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of . Similarly, in the attention relation o, express and have the same set of nodes The hypergraph of include and All simple edges and hyperedges in the graph, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of .

[0072] In order to enable the expert model to learn from different subgraphs respectively, the present invention provides a (in Can be a friendship Or follow the relationship ) Define a general gating network First, the joint user features obtained above are Input to the gating network middle.

[0073] The mathematical representation of the gating network is:

[0074] ,

[0075] in and It's a relationship The weights and bias parameters of the gating network.

[0076] Output is a three-element vector , where each Representative relationship Middle The weight of the expert model.

[0077] Therefore, for each relationship type , three different types of expert models (simple graph , Hypergraph and composite graph ) are combined together. This process is expressed as:

[0078] ,

[0079] in It's a relationship Middle Expert Model, is the weight assigned by the gating network, Representing relationships No. The type of graph that the expert model handles.

[0080] The relationship fusion network module uses a multi-head self-attention mechanism to fuse different types of relationship features for user classification.

[0081] After obtaining the user-specific vector representing the friend and follow relationship characteristics and Finally, the present invention adopts a multi-head self-attention mechanism to promote the interaction between these two types of relationship representations. This design aims to fuse social network information and user behavior features. Specifically, the features are weighted and merged through the self-attention mechanism to enhance their complementarity. The fusion operation is defined as:

[0082] ,

[0083] in represents a multi-head self-attention layer, and They are the friend and follow relationship features after applying attention interaction.

[0084] In addition, the present invention utilizes an attention-based approach to evaluate the consistency of user representations and the synergy between different social relationship representations.

[0085] The generation of the consistency vector can be expressed as:

[0086] .

[0087] is the consistency vector, represents the attention matrix generated by the self-attention module, Represents the 2D convolution operation performed by the filter for feature extraction. The matrix is ​​reduced in dimension by the maximum pooling downsampling method, and then 2D convolution operation is performed with the filter to further extract features; then, the output is flattened into a consistency vector .

[0088] The feature fusion layer simultaneously learns the friend relationships and follow relationships associated with the user to classify the user as a social robot or a real user.

[0089] The feature fusion layer is used to integrate the friend relationship features , attention relationship characteristics and consistency vector splicing, through the output layer and The layer is transformed and expressed as:

[0090] ,

[0091] in It is for users The model prediction results are and represents a learnable parameter.

[0092] The present invention uses the cross entropy loss function and the regularization term to train the model, and the loss function is expressed as:

[0093] ,

[0094] in is the set of annotated users, represents the i-th user in the user set, is the true label, represents all trainable parameters in the model, is a hyperparameter. Represents training parameters.

[0095] The social robot detection method for false information in the electricity market according to an embodiment of the present invention includes a construction step and a detection step.

[0096] Construction steps: Build and train the HyperBot model of a social robot used to publish false fault information in the power market. The HyperBot model includes a hypergraph construction module, a multimodal user network module, a hypergraph expert network module, a relationship fusion network module, and a feature fusion layer.

[0097] The hypergraph construction module constructs a hypergraph model, uses the hypergraph model to build a social network, and captures the friendship and attention relationships of users in the social network;

[0098] The multimodal user network module extracts and fuses features from the user's text content and metadata to obtain a multimodal representation of each user;

[0099] The hypergraph expert network module learns user relationships in simple graphs, hypergraphs, and composite graphs of simple graphs and hypergraphs;

[0100] The relationship fusion network module uses a multi-head self-attention mechanism to fuse different types of relationship features for user classification;

[0101] The feature fusion layer simultaneously learns the friend relationships and follow relationships related to the user, and classifies the user as a social robot or a real user;

[0102] Detection steps: Input the user information of the user who publishes the fault information in the power market into the HyperBot model, and use the HyperBot model to determine whether the user is a social robot or a real user.

[0103] As an implementation method, the multimodal user network module includes a description encoder, a tweet encoder, a category encoder, and a value encoder.

[0104] The description encoder uses a pre-trained language model to encode user descriptions and extract user description features;

[0105] The tweet encoder uses a pre-trained language model to encode user tweets and extract the features of the user's tweets;

[0106] The category encoder classifies the user profile, uses one-hot encoding for the features of each category, and concatenates the encoded category features to obtain joint category features. The joint category features are learned by attaching a multi-layer perceptron to obtain category features related to the user.

[0107] The numerical encoder obtains the numerical characteristics of the user;

[0108] The multimodal user network module encodes and connects the description features, tweet features, category features and numerical features to obtain the joint user features of each user.

[0109] As an implementation method, the hypergraph expert network module includes multiple layers of hypergraph convolution and hybrid expert structure.

[0110] The multi-layer hypergraph convolution consists of L layers of hypergraph convolution layers, and the hypergraph convolution layer is expressed as:

[0111] ;

[0112] in It is The signal of the hypergraph convolutional layer, , H is the hypergraph incidence matrix, represents the transposed matrix of the hypergraph incidence matrix, W represents the weight of the hyperedge, and denote the diagonal matrices used for normalization of vertex and edge degrees, express The corresponding inverse matrix is, express The corresponding negative half-power matrix, represents a nonlinear activation function, Indicates The learnable filter matrix of the hypergraph convolutional layer;

[0113] The hybrid expert structure consists of two parts: , Hypergraph and composite graph There are three expert models for learning, among which,

[0114] The combined graph of friend relationship f and follow relationship o is defined as:

[0115] ;

[0116] For a friendship relationship f, Representation and and have the same set of nodes The hypergraph of include and All simple edges and hyperedges in the graph, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of ;

[0117] Similarly, in the attention relationship o, express and have the same set of nodes The hypergraph of include and All simple edges and hyperedges in the graph, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of ;

[0118] Define a general gating network for each relationship type r , r is the friend relationship f or the attention relationship o, and the joint user feature X obtained by the multimodal user network module is input into the gating network middle,

[0119] Gated Network It is expressed as:

[0120] ;

[0121] in and Is the relationship type The weights and bias parameters of the gating network, The output is a three-element vector , where each Represents the relationship type Middle The weight of the expert model;

[0122] For each relationship type r, the outputs of the three expert models are combined together, and the combination process is expressed as:

[0123] ;

[0124] in It's a relationship No. Expert models, Representing relationships No. The type of graph that the expert model handles.

[0125] As an implementation method, the construction step is followed by a training step:

[0126] The HyperBot model is trained using the cross entropy loss function and regularization term. The loss function is expressed as:

[0127] ;

[0128] in is the set of annotated users, represents a user in the user set, Is a user The real label, Represents all trainable parameters in the HyperBot model, is a hyperparameter, Represents training parameters.

[0129] In order to evaluate the performance of the present invention, the present invention conducts experiments on three real-world social robot detection datasets. Experimental data introduction:

[0130] Cresci-15, TwiBot-20, and TwiBot-22. These datasets are standard benchmarks for social robot detection and provide a variety of relationships that support graph-based models. We used the same data partitioning provided in the benchmarks to ensure that our results can be directly compared with previous studies. Table 1 shows the specific statistics of these datasets.

[0131] Table 1

[0132]

[0133] Experimental environment and parameter settings:

[0134] The relevant experiments are run on the server, and the environment configuration used is shown in Table 2.

[0135] Table 2

[0136]

[0137] In our experiments, we use AdamW as the optimizer, and the learning rate and weight decay are set to and . Regularization parameter Set to 1e-5. The number of attention heads is set to 2. The function is used as the activation function. We use a batch size of 256 and set the maximum number of learning epochs to 300. To avoid overfitting, we use Technique and early stopping.

[0138] We compare our model with mainstream social media detection methods, including feature-based, text-based, and graph-based methods:

[0139] SGBot uses random forest classifier to perform feature engineering on user metadata and derived features.

[0140] Wei et al. used a recurrent neural network to capture features in tweets and transformed social bot detection into a text classification task.

[0141] Varol et al. adopted a random forest approach that combined metadata, network features, and other derived statistics to detect various types of Twitter bots.

[0142] Alhosseini et al. used graph convolutional networks to learn user representations and perform social robot detection by aggregating neighborhood feature information.

[0143] SATAR builds a self-supervised representation learning framework that integrates user tweets, metadata, and neighborhood information, and fine-tunes it on a specific bot detection dataset to improve accuracy.

[0144] Botometer uses over 1,000 features to assess the similarity of Twitter accounts to known social bots.

[0145] BotRGCN constructs a heterogeneous graph to represent the user network and utilizes relational graph convolutional networks for representation learning and bot detection.

[0146] RGT uses friend and follow relationships to build a relational graph transformer, modeling heterogeneous influences between users and learning node representations.

[0147] We evaluate our method and baselines on three Twitter bot detection benchmarks, as shown in Table 3. The experimental results are slightly different from the original paper, which we believe may be due to the changes in the experimental settings.

[0148] Our method significantly outperforms the baseline models on all three datasets. Notably, our method achieves 1.44% higher accuracy and 1.25% higher F1 score on TwiBot-20 compared to the state-of-the-art method RGT. Moreover, it achieves 1.32% higher accuracy and 2.39% higher F1 score on TwiBot-22. HyperBot shows the best performance on the datasets collected in 2015, 2020, and 2022, demonstrating its generalization ability for detecting social robots in the real world.

[0149] Graph-based methods such as HyperBot and BotRGCN generally outperform feature-based and text-based methods such as SGBot and Wei et al. This highlights the important role that social relationships play in detecting social robots and emphasizes the necessity of building social networks for detection purposes.

[0150] HyperBot outperforms other methods using simple graphs on the Cresci-15, TwiBot-20, and TwiBot-22 datasets, indicating that it more effectively utilizes the content in social networks and is able to mine richer information from social relationships.

[0151] Table 3

[0152]

[0153] Social media bot detection is an important and challenging task. This paper uses hypergraphs to construct social networks, and adopts hypergraph expert networks and relation fusion networks for representation learning and bot detection. HyperBot aims to use the relationships between multiple users to help detect bots in social networks. We conducted extensive experiments and the bot detection based on HyperBot consistently outperformed other competitive baselines. Other studies further demonstrated the effectiveness of our model in mining the potential relationships of multiple users, indicating that our model can help in the task of social media bot detection.

[0154] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A social robot detection system for false information in the power market, characterized by: The method comprises a construction unit and a detection unit, wherein the construction unit constructs a HyperBot model of a social robot for publishing false fault information in the power market, and the detection unit inputs user information of the social robot for publishing fault information in the power market into the HyperBot model, and determines whether the user is a social robot or a real user through the HyperBot model; The HyperBot model includes a hypergraph construction module, a multimodal user network module, a hypergraph expert network module, a relationship fusion network module, and a feature fusion layer. The hypergraph construction module constructs a hypergraph model, uses the hypergraph model to build a social network, and captures the friendship and attention relationships of users in the social network; The multimodal user network module extracts and fuses features from the user's text content and metadata to obtain a multimodal representation of each user; The hypergraph expert network module learns user relationships in simple graphs, hypergraphs, and composite graphs of simple graphs and hypergraphs; The relationship fusion network module uses a multi-head self-attention mechanism to fuse different types of relationship features for user classification; The feature fusion layer simultaneously learns the friend relationships and follow relationships related to the user, and classifies the user as a social robot or a real user; The multimodal user network module includes a description encoder, a tweet encoder, a category encoder, and a value encoder. The description encoder uses a pre-trained language model to encode user descriptions and extract user description features; The tweet encoder uses a pre-trained language model to encode user tweets and extract the features of the user's tweets; The category encoder classifies the user profile, uses one-hot encoding for the features of each category, and concatenates the encoded category features to obtain joint category features. The joint category features are learned by attaching a multi-layer perceptron to obtain category features related to the user. The numerical encoder obtains the numerical characteristics of the user; The multimodal user network module encodes and connects the description features, tweet features, category features and numerical features to obtain the joint user features of each user.

2. The social robot detection system for false information in the power market as claimed in claim 1, characterized in that: The hypergraph expert network module contains multi-layer hypergraph convolution and hybrid expert structure. The multi-layer hypergraph convolution consists of L layers of hypergraph convolution layers, and the hypergraph convolution layer is expressed as: ； in It is The signal of the hypergraph convolutional layer, , H is the hypergraph incidence matrix, represents the transposed matrix of the hypergraph incidence matrix, W represents the weight of the hyperedge, and denote the diagonal matrices used for normalization of vertex and edge degrees, express The corresponding inverse matrix is, express The corresponding negative half-power matrix, represents a nonlinear activation function, Indicates The learnable filter matrix of the hypergraph convolutional layer; The hybrid expert structure consists of two parts: , Hypergraph and composite graph There are three expert models for learning, among which, The combined graph of friend relationship f and follow relationship o is defined as: ； For a friendship relationship f, Representation and and have the same set of nodes The hypergraph of include and All simple edges and hyperedges in the graph, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of ; Similarly, in the attention relationship o, express and have the same set of nodes The hypergraph of include and All simple edges and hyperedges in the graph, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of ; Define a general gating network for each relationship type r , r is the friend relationship f or the attention relationship o, and the joint user feature X obtained by the multimodal user network module is input into the gating network middle, Gated Network It is expressed as: ； in and Is the relationship type The weights and bias parameters of the gating network, The output is a three-element vector , where each Represents the relationship type Middle The weight of the expert model; For each relationship type r, the outputs of the three expert models are combined together, and the combination process is expressed as: ； in It's a relationship No. Expert models, Representing relationships No. The type of graph that the expert model handles.

3. The social robot detection system for false information in the power market as claimed in claim 2, characterized in that: The fusion operation of the relational fusion network module is defined as: ； in represents a multi-head self-attention layer, and are the friend relationship feature and attention relationship feature vectors respectively, and They are the friend relationship features and attention relationship features after applying attention interaction; The relationship fusion network module uses an attention-based approach to evaluate the consistency of user representations and the synergy between different social relationship representations. The generation of the consistency vector is expressed as: ； in, is the consistency vector, represents the 2D convolution operation performed by the filter, and M represents the attention matrix generated by the self-attention module.

4. The social robot detection system for false information in the power market as claimed in claim 3, characterized in that: The feature fusion layer combines the friend relationship features , attention relationship characteristics and consistency vector splicing, through the output layer and Layers to convert: ； in It is for users The model prediction results are and represents the learnable parameters, , , For users ’s friend relationship features, follow relationship features, and consistency vector.

5. The social robot detection system for false information in the power market as claimed in claim 1, characterized in that: The HyperBot model is trained using the cross entropy loss function and regularization term. The loss function is expressed as: ； in is the set of annotated users, represents a user in the user set, Is a user The real label, represents the training parameters, Represents all trainable parameters in the HyperBot model, is a hyperparameter.

6. A social robot detection method for false information in the power market, characterized in that: Including construction steps and detection steps, Construction steps: Construct a HyperBot model of a social robot used to publish false fault information in the power market; the HyperBot model includes a hypergraph construction module, a multimodal user network module, a hypergraph expert network module, a relationship fusion network module, and a feature fusion layer, among which, The hypergraph construction module constructs a hypergraph model, uses the hypergraph model to build a social network, and captures the friendship and attention relationships of users in the social network; The multimodal user network module extracts and fuses features from the user's text content and metadata to obtain a multimodal representation of each user; The hypergraph expert network module learns user relationships in simple graphs, hypergraphs, and composite graphs of simple graphs and hypergraphs; The relationship fusion network module uses a multi-head self-attention mechanism to fuse different types of relationship features for user classification; The feature fusion layer simultaneously learns the friend relationships and follow relationships related to the user, and classifies the user as a social robot or a real user; Detection step: Input the user information of the user who publishes the fault information in the power market into the HyperBot model, and use the HyperBot model to determine whether the user is a social robot or a real user; The multimodal user network module includes a description encoder, a tweet encoder, a category encoder, and a value encoder. The description encoder uses a pre-trained language model to encode user descriptions and extract user description features; The tweet encoder uses a pre-trained language model to encode user tweets and extract the features of the user's tweets; The category encoder classifies the user profile, uses one-hot encoding for the features of each category, and concatenates the encoded category features to obtain joint category features. The joint category features are learned by attaching a multi-layer perceptron to obtain category features related to the user. The numerical encoder obtains the numerical characteristics of the user; The multimodal user network module encodes and connects the description features, tweet features, category features and numerical features to obtain the joint user features of each user.

7. The social robot detection method for false information in the power market as claimed in claim 6, characterized in that: The hypergraph expert network module contains multi-layer hypergraph convolution and hybrid expert structure. The multi-layer hypergraph convolution consists of L layers of hypergraph convolution layers, and the hypergraph convolution layer is expressed as: ； in It is The signal of the hypergraph convolutional layer, , H is the hypergraph incidence matrix, represents the transposed matrix of the hypergraph incidence matrix, W represents the weight of the hyperedge, and denote the diagonal matrices used for normalization of vertex and edge degrees, express The corresponding inverse matrix is, express The corresponding negative half-power matrix, represents a nonlinear activation function, Indicates The learnable filter matrix of the hypergraph convolutional layer; The hybrid expert structure consists of two parts: , Hypergraph and composite graph There are three expert models for learning, among which, The combined graph of friend relationship f and follow relationship o is defined as: ； For a friendship relationship f, Representation and and have the same set of nodes The hypergraph of include and All simple edges and hyperedges in the graph, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of ; Similarly, in the attention relationship o, express and have the same set of nodes The hypergraph of include and All simple edges and hyperedges in the graph, and Indicates the merger from and The hypergraph of all simple edges and hyperedges of ; Define a general gating network for each relationship type r , r is the friend relationship f or the attention relationship o, and the joint user feature X obtained by the multimodal user network module is input into the gating network middle, Gated Network It is expressed as: ； in and Is the relationship type The weights and bias parameters of the gating network, The output is a three-element vector , where each Represents the relationship type Middle The weight of the expert model; For each relationship type r, the outputs of the three expert models are combined together, and the combination process is expressed as: ； in It's a relationship No. Expert models, Representing relationships No. The type of graph that the expert model handles.

8. The social robot detection method for false information in the power market as claimed in claim 6, characterized in that: The build step is followed by the training step: The HyperBot model is trained using the cross entropy loss function and regularization term. The loss function is expressed as: ； in is the set of annotated users, represents a user in the user set, Is a user The real label, Represents all trainable parameters in the HyperBot model, is a hyperparameter, Represents training parameters.

Images