A social robot user detection model based on emotion-behavior collaborative analysis
By combining the DistilBERT and Louvain algorithms with the Light-HGNN social machine user detection model, the problems of machine user emotion and behavior spoofing and sample imbalance were solved, achieving high-precision and robust machine user recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHONGQING UNIV OF POSTS & TELECOMM
- Filing Date
- 2026-05-08
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies struggle to effectively distinguish between machine users and human users, especially when disguised by emotional expression and coordinated group behavior. Furthermore, imbalanced sample sizes can lead to false positives, compromising the security of social networks.
The DistilBERT model is used to analyze the subtle fluctuations in text sentiment. The Louvain algorithm is combined to construct a hypergraph and Light-HGNN is used to deeply mine group collaborative behavior. The feature distribution is reconstructed through NearMiss-3 and adversarial SMOTE techniques. The training process is optimized by embedding a dynamic weighted cross-entropy loss function to improve the robustness and accuracy of the model.
It significantly improves the detection accuracy and robustness for machine users, accurately identifies spoofing behavior and quantity imbalance issues, and provides decision-making transparency.
Smart Images
Figure CN122490264A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of social network user detection, specifically involving a social machine user detection model based on sentiment-behavior collaborative analysis. Background Technology
[0002] In the digital age of highly convenient information dissemination, bot users on social media, through comments and forwarding of false information or rumors, may cause various risks such as fraudulent marketing and negative public opinion, potentially weakening users' rational judgment. Therefore, accurately and quickly determining whether a user is a bot based on their behavioral patterns, semantic features, and social network structure is particularly important. In recent years, bot detection has attracted a large number of experts and scholars to conduct research. Depending on their research focus, these research methods can be broadly divided into two directions: one is detection technology based on pre-trained language models and multimodal fusion, which improves the detection of bots by analyzing user text, images, and other modal data; the other is detection methods based on graph neural networks and structured relationship modeling, which model user social relationships and distinguish the differences in interaction between users and bots.
[0003] With the rapid development of artificial intelligence, significant success has been achieved in identifying machine users on social networks. However, research and analysis of machine user groups under emotional and collaborative behaviors still face several challenges:
[0004] 1. The similarity between machine users and human emotional expression. In social networks, machine users exhibit a similarity to human emotional expression. Using natural language processing technology, machine users employ emotional vocabulary, emoticons, and specific sentence structures to simulate human emotional expression. This high degree of realism easily confuses traditional detection methods.
[0005] 2. The problem of masquerading collaborative behavior among machine users. Machine users can simulate the natural behavior patterns of real user groups in terms of time, frequency, and content through synchronized operations and decentralized interactions. This collaborative masquerading effectively avoids detection methods that rely on individual anomalies or single behavioral characteristics.
[0006] 3. The imbalance between the number of machine users and real users. Due to the complexity of machine user behavior patterns and the variety of generation methods, obtaining samples from machine users is difficult and the number is limited. In contrast, the number of real users is vast and samples are readily available. This imbalance can easily lead to model misjudgments. Summary of the Invention
[0007] To address the aforementioned technical problems, this invention provides a social machine user detection model based on sentiment-behavior collaborative analysis. It includes the following steps:
[0008] S1. Use the DistilBERT two-layer architecture to analyze text sentiment micro-fluctuations and extract user sentiment micro-fluctuation features;
[0009] S2. Use the Louvain algorithm for topological preprocessing to construct a hypergraph, and then use Light-HGNN to deeply mine group cooperative behavior.
[0010] S3. It uses NearMiss-3 and adversarial SMOTE techniques to reconstruct the semantic and behavioral feature distribution, embeds a dynamic weighted cross-entropy loss function to optimize the training process, solves the sample imbalance problem, improves the robustness of the model, and achieves accurate identification of machine users.
[0011] The beneficial effects of this invention are:
[0012] This invention introduces the DistilBERT model to extract high-dimensional text semantics and incorporates an emotional entropy temporal attention mechanism. By quantifying the evolutionary inertia of emotional polarity and the alignment with contextual logic, it accurately captures the emotional discontinuities and context-detached features exposed by machine users during the masquerade process. This invention also provides a community-temporal two-layer cascaded group collaborative detection method. It uses the Louvain algorithm to quickly locate suspicious high-density communities based on the original interaction topology, and leverages Light-HGNN to mine behavioral associations and operational temporal synchronization between nodes within the community, effectively identifying the dispersed masquerade strategies of machine user groups. Furthermore, this invention provides a category balancing strategy based on dynamic weighting and adversarial reconstruction. It combines NearMiss-3 and adversarial SMOTE to reconstruct the sample distribution in the multimodal feature space, and introduces SHAP attribution analysis to alleviate data imbalance problems, significantly improving the detection accuracy and robustness of machine users. Attached Figure Description
[0013] Figure 1 This is a schematic diagram of the process of the present invention;
[0014] Figure 2 This is a schematic diagram illustrating the extraction of subtle emotional fluctuations.
[0015] Figure 3 A schematic diagram of user group collaborative feature extraction;
[0016] Figure 4 This is a schematic diagram of machine user detection based on adversarial equilibrium. Detailed Implementation
[0017] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0018] A social machine user detection model based on sentiment-behavior collaborative analysis, the process of which is as follows: Figure 1 As shown, the specific steps include:
[0019] S1. Use the DistilBERT two-layer architecture to analyze text sentiment micro-fluctuations and extract user sentiment micro-fluctuation features;
[0020] S2. Use the Louvain algorithm for topological preprocessing to construct a hypergraph, and then use Light-HGNN to deeply mine group cooperative behavior.
[0021] S3. It uses NearMiss-3 and adversarial SMOTE techniques to reconstruct the semantic and behavioral feature distribution, embeds a dynamic weighted cross-entropy loss function to optimize the training process, solves the sample imbalance problem, improves the robustness of the model, and achieves accurate identification of machine users.
[0022] By using the DistilBERT two-layer architecture to analyze text sentiment micro-fluctuations and extract user sentiment micro-fluctuation features, such as... Figure 2 As shown, it specifically includes:
[0023] S11, Given a user's historical text sequence ,in Indicates the first Each text item Contains word sequence , Given the text length, the input text is processed through the embedding layer of a pre-trained DistilBERT. :
[0024]
[0025] in, For word embedding matrix, For position encoding, , For the hidden layer dimension.
[0026] S12. Using the sine function to generate position vectors The word embeddings are added to the positional encodings to obtain the initial hidden state matrix. The initial hidden state matrix The input consists of a stack of six Transformer encoder layers in the DistilBERT model. Each layer contains a multi-head self-attention network and a feedforward network, and residual connections and layer normalization are used. Layer output :
[0027]
[0028] S13. Take the vector at the first position of the output of the last layer of encoding as the semantic representation of the entire text sequence. Then, the semantic vectors are mapped to probability distributions of discrete sentiment categories to capture the sentiment expression tendency of the text. This is achieved by constructing a trainable sentiment classifier, which maps the semantic vectors to sentiment probability distributions through trainable linear layers and a softmax function. :
[0029]
[0030] in, This is the weight matrix. For the bias term, calculate the probability distribution. Each dimension reflects the intensity of a user's expression of a specific emotion.
[0031] S14, Emotional Entropy The feature vector is mapped to the same dimension as the semantic vector and injected into the semantic space through residual connections, resulting in a feature vector that integrates the original semantic information and sentiment entropy information. :
[0032]
[0033] in, For projection weights. This is a learnable bias term.
[0034] S15. By injecting emotional uncertainty features into the semantic space through residual connections, fused features are obtained. :
[0035]
[0036] in, This is a learnable modulating factor. The formula enables the model to dynamically adjust the distribution of semantic features based on the degree of emotional confusion: if the machine user's emotional logic is abnormal, this layer will significantly alter the original semantic representation.
[0037] S16. Capture anomalous shifts in emotion over time, explicitly embed attention weights to obtain intermediate hidden layer vectors. :
[0038]
[0039] in, , This is the weight matrix. For the intensity of emotional changes, This is a bias term.
[0040] S17. Capture anomalous shifts in emotion over time, introduce a temporal attention mechanism to aggregate all historical text features of the user, and calculate an attention score. :
[0041]
[0042] in, This is the context query vector.
[0043] S18. Obtain the user's sentiment semantic comprehensive feature vector through weighted aggregation. :
[0044]
[0045] The vector It deeply integrates the user's global semantic information and local emotional evolution anomaly features, which will serve as an important input for subsequent modules.
[0046] The Louvain algorithm is used for topology preprocessing to construct a hypergraph, and Light-HGNN is used for deep mining of group cooperative behaviors, such as... Figure 3 As shown, it specifically includes:
[0047] S21. Extract user characteristic behaviors, construct a weighted heterogeneous graph of user behavior features, and base the graph on the user behavior collaboration index. To construct a weighted neighbor matrix :
[0048]
[0049] in, This is a normalized value representing the number of direct interactions between users. The time window length, This is the balance coefficient.
[0050] S22, Transform the heterogeneous adjacency matrix This serves as the basic input for the Louvain algorithm. After constructing the cardinality-dominance adjacency matrix, the user interaction density is calculated. :
[0051]
[0052] in, For nodes The degree, For the total weight, when hour And explain the user and Within the same community. Through iterative optimization. The value divides the entire image into... A community collection .
[0053] S23. Use hypergraph neural networks to process community structure and construct hypergraph association matrices. If the user Belongs to the community ,but Based on this, Light-HGNN
[29] is introduced for feature propagation. In order to prevent the deep propagation from being too smooth, Light-HGNN removes the non-linear activation function, and its core propagation formula is as follows:
[0054]
[0055] in, Representative at the The node feature matrix of the layer and These are diagonal matrices representing node degree and hyperedge degree, respectively. This formula, through hyperedge relay, achieves global aggregation and smoothing of node features within the community.
[0056] S24, after After layer propagation, a hierarchical weighted mechanism is used to fuse neighborhood information with different hop counts to obtain the final structural embedding representation of the node. :
[0057]
[0058] in, Indicates the first Layer weight coefficients.
[0059] S25. Further capture the operational synchronicity of the machine group in the time dimension and calculate the average absolute deviation within the community. And map it back to the node:
[0060]
[0061] in, These are community units detected by the Lovain algorithm. This represents the average of the time characteristics of all users within the community. Timestamp of a specific operation. Abnormally low values can reveal the automatic synchronization behavior of user clusters.
[0062] S26. Concatenate the structural embeddings extracted by Light-HGNN with the temporal synchronization features to obtain the group cooperative behavior vector. :
[0063]
[0064] The vector Deep coding reveals the group collaboration patterns of users in social networks, providing crucial evidence for identifying covert group bot users.
[0065] The semantic and behavioral feature distributions are reconstructed using NearMiss-3 and adversarial SMOTE techniques, and the training process is optimized by embedding a dynamically weighted cross-entropy loss function. Figure 4 As shown, it specifically includes:
[0066] S31. Using the emotion vector obtained in claim 2 and the behavior vector obtained by claim 3 and coded user basic attribute features By concatenating the features, a comprehensive feature vector is obtained. .
[0067] S32, For the majority class sample set (Real users) The NearMiss-3 algorithm is introduced to filter high-value samples located at the classification boundary. Calculate the value of each majority class sample. To the minority class set The closest in the middle The average Euclidean distance of the samples:
[0068]
[0069] in, Represents the distance in the user set No. Closest samples, Euclidean distance. This is a hyperparameter. According to... Sort the majority class samples in ascending order and retain the samples with the smallest distance. This step effectively preserves "hard-to-classify samples" in areas overlapping with the distribution of machine users, forcing the model to focus on the classification boundary while removing redundant data far from the decision surface.
[0070] S33. For each minority class sample The SMOTE oversampling technique is introduced to randomly select its k nearest neighbor samples in the feature space. Interpolation coefficients are generated according to a uniform distribution to synthesize intermediate samples. :
[0071]
[0072] Meanwhile, for each intermediate sample generated by SMOTE Adding adversarial perturbations provides the option to maximize the current classifier. The loss function is calculated, and the perturbation is determined. :
[0073]
[0074] in, This is the disturbance factor, used to control the strength of the noise suppression. The loss function is relative to the synthetic sample. The gradient.
[0075] S34, Through The adversarial samples are added to the original synthetic samples to obtain adversarial samples, forcing the model to learn more robust decision boundaries. Then, class balance is calculated, assuming the current training batch... Includes A majority class sample and Calculate the current balance coefficient for a minority class sample. :
[0076]
[0077] in, After obtaining the current balance coefficient through the training steps, dynamic weight allocation is performed.
[0078] S35, For each category Calculate the dynamic weights of its minority and majority classes. :
[0079]
[0080] in, As a smoothing factor, it prevents the denominator from being zero; this weight... This will directly affect the subsequent loss function calculation, enabling the model to maintain gradient sensitivity to minority class samples even in batches dominated by the majority class.
[0081] After obtaining the reconstructed sample features and dynamic weight parameters, a dual-hidden-layer MLP architecture is used for deep feature interaction and classification decision-making. During training, the model uses a combined dynamic weighting mechanism and backpropagation algorithm for parameter iteration to ensure sensitivity to machine-user samples, including:
[0082] S351, Resample the feature vector The input to the first hidden layer undergoes a linear transformation through a fully connected layer, and the ReLU activation function is used to extract the non-linear interaction relationships between features:
[0083]
[0084] in and These are the weight matrix and bias terms for the first layer, respectively. The output features of the first layer. S352, Introduce Dropout regularization mechanism in the second hidden layer:
[0085]
[0086] in, and For the second layer parameters, This represents the probability of random discarding.
[0087] S353. After deep feature extraction, the features are mapped to the real number domain through the output layer, and the Sigmoid function is used to calculate the predicted probability that the user belongs to the machine user. :
[0088]
[0089] S354, Combining dynamic weights Construct a weighted binary cross-entropy loss function to obtain the dynamic weighted loss. :
[0090]
[0091] in, For batch size, This is a real label.
[0092] S355, Using the AdamW optimizer to optimize network parameters Iterative updates are performed to achieve synergistic convergence of feature extraction and classification decisions.
[0093] S36. Calculate each feature SHAP value Quantify the expected marginal contribution of this feature to the model's prediction results:
[0094]
[0095] in, For the complete set of features, For feature subset, For using subsets The model prediction results.
[0096] S37. Calculate the contribution value of each feature. Set its threshold, when the contribution value If the value exceeds its threshold, it is selected as a key feature, and the calculation is performed. Then, the mean global absolute SHAP value is calculated for each feature. :
[0097]
[0098] S38, Press By sorting in descending order, we can finally clarify the specific contribution weights of features such as emotional evolution inertia and synergy index in determining whether a user is a machine user, thus giving the model decision-making transparency in the field of security detection.
[0099] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.
Claims
1. A social machine user detection model based on sentiment-behavior collaborative analysis, characterized in that, include: S1. Use the DistilBERT two-layer architecture to analyze text sentiment micro-fluctuations and extract user sentiment micro-fluctuation features; S2. Use the Louvain algorithm for topological preprocessing to construct a hypergraph, and then use Light-HGNN to deeply mine group cooperative behavior. S3. It uses NearMiss-3 and adversarial SMOTE techniques to reconstruct the semantic and behavioral feature distribution, embeds a dynamic weighted cross-entropy loss function to optimize the training process, solves the sample imbalance problem, improves the robustness of the model, and achieves accurate identification of machine users.
2. The social machine user detection model based on sentiment-behavior co-analysis according to claim 1, characterized in that, By using the DistilBERT two-layer architecture to analyze text sentiment micro-fluctuations, user sentiment micro-fluctuation features are extracted, including: S11, Given a user's historical text sequence ,in Indicates the first Each text item Contains word sequence , Given the text length, the input text is processed through the embedding layer of a pre-trained DistilBERT. : ； in, For word embedding matrix, For position encoding, , For the hidden layer dimension; S12. Using the sine function to generate position vectors The word embeddings are added to the positional encodings to obtain the initial hidden state matrix. ; Set the initial hidden state matrix The input consists of a stack of 6 Transformer encoder layers in the DistilBERT model. Each layer contains a multi-head self-attention network and a feedforward network, and residual connections and layer normalization are used. Layer output : ； S13. Take the vector at the first position of the output of the last layer of encoding as the semantic representation of the entire text sequence. Then, the semantic vectors are mapped to probability distributions of discrete sentiment categories to capture the sentiment expression tendency of the text. This is achieved by constructing a trainable sentiment classifier, which maps the semantic vectors to sentiment probability distributions through trainable linear layers and a softmax function. : ； in, This is the weight matrix. For the bias term, calculate the probability distribution. Each dimension reflects the intensity of a user's expression of a specific emotion; S14, Emotional Entropy The feature vector is mapped to the same dimension as the semantic vector and injected into the semantic space through residual connections, resulting in a feature vector that integrates the original semantic information and sentiment entropy information. : ； in, For projection weights; For learnable bias terms; S15. By injecting emotional uncertainty features into the semantic space through residual connections, fused features are obtained. : ； in, This is a learnable modulating factor; the formula enables the model to dynamically adjust the distribution of semantic features according to the degree of emotional confusion: if the machine user's emotional logic is abnormal, this layer will significantly change the original semantic representation. S16. Capture anomalous shifts in emotion over time, explicitly embed attention weights to obtain intermediate hidden layer vectors. : ； in, , This is the weight matrix. For the intensity of emotional changes, For bias terms; S17. Capture anomalous shifts in emotion over time, introduce a temporal attention mechanism to aggregate all historical text features of the user, and calculate an attention score. : ； in, For context query vector; S18. Obtain the user's sentiment semantic comprehensive feature vector through weighted aggregation. ; ； The vector It deeply integrates the user's global semantic information and local emotional evolution anomaly features, which will serve as an important input for subsequent modules.
3. The social machine user detection model based on sentiment-behavior co-analysis according to claim 1, characterized in that, The Louvain algorithm is used for topology preprocessing to construct a hypergraph, and Light-HGNN is used for deep mining of community cooperative behavior, including: S21. Extract user characteristic behaviors, construct a weighted heterogeneous graph of user behavior features, and base the graph on the user behavior collaboration index. To construct a weighted neighbor matrix : ； in, This is a normalized value representing the number of direct interactions between users. The time window length, This is the balance coefficient; S22, Transform the heterogeneous adjacency matrix As the basic input to the Louvain algorithm; after constructing the cardinal-to-dominant adjacency matrix, the user interaction density is calculated. : ； in, For nodes The degree, For the total weight, when hour And explain the user and Within the same community; through iterative optimization The value divides the entire image into... A community collection ; S23. Use hypergraph neural networks to process community structure and construct hypergraph association matrices. If the user Belongs to the community ,but Based on this, Light-HGNN[29] is introduced for feature propagation; in order to prevent the deep propagation from being too smooth, Light-HGNN removes the non-linear activation function, and its core propagation formula is as follows: ； in, Representative at the The node feature matrix of the layer and These are diagonal matrices representing node degree and hyperedge degree, respectively; this formula achieves global aggregation and smoothing of node features within the community through hyperedges. S24, after After layer propagation, a hierarchical weighted mechanism is used to fuse neighborhood information with different hop counts to obtain the final structural embedding representation of the node. : ； in, Indicates the first Layer weight coefficients; S25. Further capture the operational synchronicity of the machine group in the time dimension and calculate the average absolute deviation within the community. And map it back to the node: ； in, These are community units detected by the Lovain algorithm. This represents the average of the time characteristics of all users within the community; Timestamp of a specific operation; Abnormally low values can reveal the automatic synchronization behavior of user clusters; S26. Concatenate the structural embeddings extracted by Light-HGNN with the temporal synchronization features to obtain the group cooperative behavior vector. : ； The vector Deep coding reveals the group collaboration patterns of users in social networks, providing crucial evidence for identifying covert group bot users.
4. The social machine user detection model based on sentiment-behavior co-analysis according to claim 1, characterized in that, The semantic and behavioral feature distributions are reconstructed using NearMiss-3 and adversarial SMOTE techniques, and the training process is optimized by embedding a dynamically weighted cross-entropy loss function, including: S31. Using the emotion vector obtained in claim 2 and the behavior vector obtained by claim 3 and coded user basic attribute features By concatenating the features, a comprehensive feature vector is obtained. ; S32, For the majority class sample set (Real users) The NearMiss-3 algorithm is introduced to filter high-value samples located at the classification boundary; the calculation of each majority class sample To the minority class set The closest in the middle The average Euclidean distance of the samples: ； in, Represents the distance in the user set No. Closest samples, Euclidean distance. For hyperparameters; according to Sort the majority class samples in ascending order and retain the samples with the smallest distance. One sample; S33. For each minority class sample The SMOTE oversampling technique is introduced to randomly select its k nearest neighbor samples in the feature space. Interpolation coefficients are generated according to a uniform distribution to synthesize intermediate samples. : ； Meanwhile, for each intermediate sample generated by SMOTE Adding adversarial perturbations provides the option to maximize the current classifier. The loss function is calculated, and the perturbation is determined. : ； in, This is the disturbance factor, used to control the strength of the noise suppression. The loss function is relative to the synthetic sample. The gradient; S34, Through The adversarial samples are added to the original synthetic samples to obtain adversarial samples, forcing the model to learn more robust decision boundaries; then, class balance is calculated, assuming the current training batch... Includes A majority class sample and Calculate the current balance coefficient for a minority class sample. : ； in, After obtaining the current balance coefficient for the training steps, dynamic weight allocation is performed. S35, For each category Calculate the dynamic weights of its minority and majority classes. : ； in, As a smoothing factor, it prevents the denominator from being zero; this weight... This will directly affect the subsequent loss function calculation, enabling the model to maintain gradient sensitivity to minority class samples even in batches dominated by the majority class. S36. Calculate each feature SHAP value Quantify the expected marginal contribution of this feature to the model's prediction results: ； in, For the complete set of features, For feature subset, For using subsets The model prediction results; S37. Calculate the contribution value of each feature. Set its threshold, when the contribution value If the value exceeds its threshold, it is selected as a key feature, and the calculation is performed. Then, the mean global absolute SHAP value is calculated for each feature. : ； S38, Press By sorting in descending order, we can finally clarify the specific contribution weights of features such as emotional evolution inertia and synergy index in determining whether a user is a machine user, thus giving the model decision-making transparency in the field of security detection.
5. A social machine user detection model based on sentiment-behavior co-analysis according to claim 4, characterized in that, After obtaining the reconstructed sample features and dynamic weight parameters, a dual-hidden-layer MLP architecture is used for deep feature interaction and classification decision-making. During training, the model uses a combined dynamic weighting mechanism and backpropagation algorithm to iterate parameters, ensuring sensitivity to machine user samples, including: S351, Resample the feature vector The input to the first hidden layer undergoes a linear transformation through a fully connected layer, and the ReLU activation function is used to extract the non-linear interaction relationships between features: ； in and These are the weight matrix and bias terms for the first layer, respectively. The output features of the first layer; S352, introduce Dropout regularization mechanism in the second hidden layer: ； in, and For the second layer parameters, This represents the probability of random discarding. S353. After deep feature extraction, the features are mapped to the real number domain through the output layer, and the Sigmoid function is used to calculate the predicted probability that the user belongs to the machine user. : ； S354, Combining dynamic weights Construct a weighted binary cross-entropy loss function to obtain the dynamic weighted loss. : ； in, For batch size, This is a real label; S355, Using the AdamW optimizer to optimize network parameters Iterative updates are performed to achieve synergistic convergence of feature extraction and classification decisions.