Text and social graph-based unreal information detection method

Abstract

The invention discloses an unreal information detection method based on a text and a social graph, and belongs to the field of natural language processing. According to the method, a hierarchical graph coding architecture is adopted, and the multi-scale extraction capability of propagation characteristics is enhanced through the synergistic effect of local attention and global topology perception; a dynamic gating mechanism is adopted, a feature fusion strategy is adjusted in a self-adaptive mode, and cooperative representation of heterogeneous modal features is optimized; in combination with a self-supervised contrast learning method based on graph structure enhancement, the influence of random disturbance in a propagation graph on a detection result is relieved, and the anti-noise capability of the model is improved; a modular system design is adopted, independent optimization and flexible replacement of each functional assembly are supported, and system function expansion can be conveniently carried out according to actual requirements.

Description

Technical Field

[0001] The present invention belongs to the field of natural language processing, and specifically relates to a false information detection method that integrates text semantic features with social network communication structure features, and is suitable for the automatic identification of false information on social media platforms. Background Art

[0002] Definitions of Key Terms:

[0003] BERT (Bidirectional Encoder Representations from Transformers): A pre-trained deep learning model based on the Transformer architecture that can learn context-sensitive semantic representations from large amounts of text. This model uses a bidirectional self-attention mechanism to simultaneously capture contextual information on both the left and right sides of a word, effectively improving its ability to express text features. BERT is widely used in natural language processing tasks such as text classification, named entity recognition, and question-answering systems, demonstrating its versatility and transferability.

[0004] Graph Attention Network (GAT): A graph neural network model that incorporates an attention mechanism, suitable for learning node representations in graph-structured data. This approach assigns different attention weights to each node's neighbors, enabling differentiated aggregation of local features and enhancing the model's ability to represent complex structures. GAT is trained primarily on local adjacency information, enabling effective modeling without accessing the entire graph structure. It is widely used in tasks such as social network modeling, node classification, and anomaly detection.

[0005] Graph Convolutional Network (GCN): A graph neural network architecture based on graph convolution, used to learn node features in graph structures. This method effectively extracts local structural information of nodes through neighborhood feature aggregation and nonlinear transformation. GCN extends traditional convolution operations to the graph domain, defining convolution propagation rules based on the graph's Laplacian matrix to achieve efficient propagation of features between nodes. It has demonstrated promising results in tasks such as node classification and social graph analysis.

[0006] In recent years, social media has become a crucial hub for information dissemination. According to 2024 data from Statista, a comprehensive global data database, over 5 billion users worldwide use social media, a figure projected to exceed 6 billion by 2028. However, the open nature of social media and the decentralized nature of content production also increase the risk of misinformation. The cross-platform spread of misinformation can not only distort individual decision-making and mislead public perception, but can also trigger social polarization and even severely impact public safety, economic order, and social atmosphere.

[0007] Currently, mainstream social media platforms primarily employ two types of detection mechanisms: manual review and automated detection. However, with the ever-increasing scale of social media data, especially the exponential growth of AI-generated content, traditional manual detection mechanisms face systemic bottlenecks such as delayed response times, limited coverage, and high operating costs. These shortcomings present a significant opportunity for the further development of automated detection methods.

[0008] Social context-based detection methods are an important technical approach for automated misinformation identification. These methods construct user interaction graphs and utilize graph neural network technology to uncover unusual patterns in dissemination paths. Compared to pure text analysis methods, these methods can additionally capture implicit patterns of dissemination among user groups. However, current mainstream methods still have certain technical limitations:

[0009] First, existing graph encoders mostly use a single type of network layer, which has a representation bottleneck when capturing local communication hotspots and global topological trends simultaneously, resulting in limited sensitivity to hierarchical communication features; second, the heterogeneity of text semantics and communication structure features makes it difficult for traditional static fusion strategies to adapt to the changes in feature weights at different communication stages; finally, random interference and adversarial communication strategies in social graphs can easily cause perturbations in the feature space, and existing methods lack effective noise filtering mechanisms to cope with the robustness challenges in complex communication environments.

[0010] These technical bottlenecks restrict the practical application of contextual detection methods in terms of early warning timeliness, adaptability to dynamic scenarios, and cross-platform robustness. Overcoming these limitations requires establishing a hierarchical graph feature extraction framework, developing a dynamic adaptive fusion mechanism, and constructing an anti-interference feature enhancement scheme. These are the core issues addressed by this patented technological innovation. Summary of the Invention

[0011] To address the shortcomings of existing technologies, this paper proposes a method for detecting misinformation on social media platforms based on text and social graphs. This method constructs a dual-channel heterogeneous feature learning framework, employs a pretrained language model to extract text semantic features, and employs a graph encoder based on a tandem GAT-GCN architecture to extract multi-level propagation structure features. Furthermore, a dynamic gating mechanism is designed to achieve adaptive spatial alignment of the two feature types. A node-degree-based graph augmentation strategy is introduced, combined with graph comparative learning methods to optimize feature representation.

[0012] The technical solution of the present invention is a method for detecting false information based on text and social graphs, which includes the following steps:

[0013] Step 1: Text feature extraction;

[0014] Perform multi-level cleaning preprocessing on raw social media text to eliminate interference from user identifiers and non-standard characters and construct a standardized input sequence. A pre-trained language model is used for contextual semantic encoding, and a multi-layer Transformer architecture is used to extract global semantic representations. A nonlinear projection layer is designed to reduce the dimensionality of high-dimensional features, generating discriminative text feature vectors suitable for the task of false information detection.

[0015] Step 2: Build a social graph;

[0016] A hierarchical communication graph structure is constructed based on source posts, with nodes representing user interaction relationships. The feature matrix uses statistical methods to extract the distribution of text keywords. A multimodal edge set is defined to model the influence of two-way communication, supporting three construction strategies: top-down, bottom-up, and undirected connections. Based on the dynamic fitness centrality calculation of edge patterns, a graph data structure that combines semantic associations and social interactions is formed, providing a joint representation space for subsequent graph neural networks.

[0017] Step 3: Graph structure feature extraction;

[0018] A hierarchical graph neural network architecture is constructed, achieving dual perception of local interaction patterns and global topological features through the sequential concatenation of GAT and GCN. A multi-head attention mechanism is used to dynamically calculate the association weights between nodes to capture key local information in the propagation path. Graph convolutional layers are used to establish residual connections, fusing shallow and deep features to enhance global topological modeling capabilities. An attention gating mechanism is introduced to evaluate node importance, and a robust graph embedding representation is generated through weighted aggregation. Finally, feature space dimension alignment is achieved through nonlinear transformation.

[0019] Step 4: Feature fusion and classification decision;

[0020] A dynamic gating mechanism is designed to achieve adaptive fusion of text and image features, dynamically adjusting the weight ratio of heterogeneous features through learnable parameters. A dimension splicing strategy is used to construct a joint representation vector, balancing the relevance of semantic details and propagation patterns. The fused features are mapped to the classification space through multiple layers of nonlinear transformations, and end-to-end false information detection decisions are achieved through probability distribution generation.

[0021] Step 5: Image enhancement contrastive learning;

[0022] A node-degree-guided self-supervised contrastive learning framework is constructed. This framework generates enhanced views of the graph structure features obtained in step 3 through a dual-path approach of topological perturbation and feature masking. The probability of edge deletion and feature masking is dynamically adjusted based on node influence, preserving key propagation paths and enhancing the model's adaptability to partial observations. A multi-view contrast optimization strategy is designed to maximize semantic consistency constraints, driving the graph encoder to learn enhanced, invariant feature representations. This strategy, combined with supervisory signals, achieves a synergistic improvement in classification performance and generalization ability.

[0023] Step 6: Model training and testing;

[0024] A dynamic weighted loss balancing strategy is designed, using a cosine annealing scheduling mechanism to coordinate the optimization strengths of the classification task and self-supervised contrastive learning. Experiments are conducted on a mainstream social media misinformation detection benchmark dataset, using both global accuracy and per-class F1 scores for performance evaluation. Finally, ablation experiments systematically validate the synergistic optimization effect of the hierarchical graph coding architecture, dynamic gating fusion, and contrastive learning module.

[0025] Furthermore, the specific method of step 1 is:

[0026] Step 1.1: Use a regular expression matching mechanism to remove user identifiers containing the "@" symbol to eliminate the interference of usernames in semantic parsing;

[0027] Step 1.2: Clean up consecutive spaces and trailing punctuation, retaining valid semantic characters, and perform text normalization, converting them to lowercase and removing non-ASCII characters to form the normalized input text. ASCII characters refer to English and common symbols, excluding special characters such as Chinese characters and emoticons.

[0028] Step 1.3: Input the pre-processed text into the pre-trained BERT model for contextual semantic encoding;

[0029] A classification marker [CLS] and a separation marker [SEP] are inserted at the beginning and end of the text to construct a structured sequence. The sequence length is standardized to a fixed size m through dynamic truncation or zero padding operations. After the sequence is input into the multi-layer Transformer architecture of BERT, the multi-head self-attention mechanism is used to calculate the contextual association weights between word units to generate a text semantic feature matrix. where d t is the mark symbol for the hidden layer dimension t to represent the text-related features;

[0030] Step 1.4: Extract the first row of feature vectors corresponding to the [CLS] tag As a global semantic representation; by constructing a nonlinear projection layer with trainable parameters, high-dimensional features are mapped to low-dimensional space: using the weight matrix W t and the bias term b t Perform linear transformation and apply ReLU activation function to finally generate text feature vector

[0031] Furthermore, the specific method of step 2 is:

[0032] Step 2.1: Construct a temporal propagation chain based on user interaction containing n i Node graph instance in Represents a set of nodes, including the root node corresponding to the original post and the remaining n nodes corresponding to the reply posts i -1 node; ε i Represents the edge set. Node feature matrix It is composed of the TF-IDF feature vectors of all posts. This feature uses the word frequency-inverse document frequency statistical method to extract the text keyword distribution pattern and construct a discriminative semantic space with domain adaptability, where d g is the feature dimension;

[0033] Step 2.2: Construct edge set ε i , including three build modes:

[0034] 1) Top-down mode: through directed edges e p→c Encode parent node v p To child node v c Direct reply relationship;

[0035] 2) Bottom-up mode: define the directed edge e in reverse c→p , highlight the child node v c For parent node v p Response effect;

[0036] 3) Undirected mode: using bidirectional edges Represents the interaction between nodes;

[0037] The edge sets of the three modes can be formally defined as:

[0038]

[0039] Node degree deg(v j ) is dynamically adapted according to the edge mode: in the top-down mode, the out-degree of each node is counted, i.e., deg(v j )=∑ k A i [j,k]; calculate the in-degree in the bottom-up mode, that is, deg(v j )=∑ k A i [k,j]. Among them, A i for The adjacency matrix, A i [j,k]=1 means there is an edge from node v j Points to node v k Although the undirected mode expands the edge set, it still maintains the degree value consistency of the original propagation chain, that is, the node degree is calculated based on the initial edge definition, ensuring that the degree size is consistent in the three modes.

[0040] Furthermore, the specific method of step 3 is:

[0041] Step 3.1: Node feature matrix of social communication graph First, through linear projection and layer normalization operations, the high-dimensional sparse TF-IDF features are mapped to a low-dimensional latent space, and then nonlinearly transformed through the ReLU activation function. The calculation process is as follows:

[0042] H (0) =ReLU(LayerNorm(XW p ))

[0043] Among them, H (0) Represents the feature matrix after dimensionality reduction; LayerNorm() represents the layer normalization operation; is the projection matrix, d0 is the latent space dimension;

[0044] Step 3.2: Use GAT-GCN tandem for feature enhancement: First, use GAT to capture local interaction patterns and use the multi-head attention mechanism to calculate the dynamic association weights between nodes; let the number of attention heads be k, and the node v i Aggregation features Expressed as:

[0045]

[0046] In the formula, || represents the symbol || represents the splicing operation on the feature dimension; Represents node v i The set of neighbor nodes of Represents the neighbor node v in the mth head j For node v i The attention weight of Represents node v j In the feature matrix H (0) The corresponding eigenvector in ; is the parameter matrix of the mth head; then the nonlinear transformation is introduced through the ELU activation function to generate the intermediate feature matrix H (1) =ELU(H (1) );

[0047] Step 3.3: Input the GAT output features into the GCN layer and establish a residual connection with the initial projection features:

[0048]

[0049] Among them, H (2) represents the intermediate feature matrix obtained in this step, is the normalized adjacency matrix, and They are the graph convolution parameter matrix and the skip connection parameter matrix respectively. This residual structure realizes the complementary transmission of multi-granularity information through the linear superposition of shallow and deep features;

[0050] Step 3.4: Use the attention gating mechanism to dynamically learn the node importance weights, and calculate the weight of each node through the Sigmoid function and the scoring function g(·) composed of two fully connected layers Then generate weighted graph embedding

[0051]

[0052] in, Represents node v i In the feature matrix H (2) The corresponding eigenvector in ;

[0053] Finally, feature space alignment is completed through nonlinear transformation:

[0054]

[0055] Where, and is a learnable parameter matrix, ensuring that the output dimension d is consistent with the text feature h t Stay consistent.

[0056] Furthermore, the specific method of step 4 is:

[0057] Step 4.1: Targeting text features and graph features The learnable gating parameter θ is used to generate the normalized weight coefficient α:

[0058] α=Sigmoid(θ)

[0059] This coefficient is constrained to the (0,1) interval through the Sigmoid function, dynamically adjusting the contribution ratio of the two types of features; on this basis, the original features are weighted and scaled:

[0060] h t′ =αh t ,h g′ =(1-α)h g

[0061] Among them, h t′ and h g′ They are weighted and scaled text features and graph features respectively;

[0062] Step 4.2: The weighted features are then concatenated along the dimension to construct a joint representation vector h fuse :

[0063]

[0064] Step 4.3: Classification decision;

[0065] Extract high-order nonlinear features z through the fully connected layer and ReLU activation function:

[0066]

[0067] in, and They are the weight matrix and bias term when performing linear transformation;

[0068] The false information detection task includes four label categories, and the four-category probability distribution is finally calculated using the softmax function:

[0069]

[0070] in, is the predicted label, and They are the weight matrix and bias term when performing linear transformation;

[0071] Model training uses the cross entropy loss function as the classification loss for end-to-end optimization, and the calculation formula is:

[0072]

[0073] Furthermore, the specific method of step 5 is:

[0074] Step 5.1: Based on node v j The degree of deg(v j ), define its logarithmic degree centrality C(v j ) to balance the scale differences:

[0075] C(v j )=log(deg(v j )+1)

[0076] Step 5.2: Calculate edge e based on node centrality i→j The enhanced weight s(v i ,v j ):

[0077]

[0078] In the edge deletion strategy, edge e is calculated based on the enhanced weight i→j The deletion probability p drop (e i→j ), this strategy prioritizes removing low-weight edges:

[0079]

[0080] Among them, β d is the baseline deletion probability; s max With s mean are the maximum and mean values ​​of the edge weights of the entire graph, respectively; ∈ is a smoothing factor to avoid the denominator being zero; after generating the retained edge set ε′ through Bernoulli sampling, the edge addition strategy introduces random perturbations to generate enhanced views The random perturbation method is: from the candidate edge set ε cand Filter unconnected edges in proportion β a Randomly added to ε′ to form an enhanced edge set ε″;

[0081] Step 5.3: Centrality-weighted feature masking;

[0082] Node v j The shielding probability p mask (v j ) is dynamically adjusted by log degree centrality:

[0083]

[0084] Among them, β m is the baseline shielding probability; C max with C mean are the maximum and mean values ​​of the logarithmic degree centrality of the nodes in the entire graph, respectively. Then, a mask vector M is generated based on the masking probability, and a masking operation is performed on the feature matrix X:

[0085] X′=X⊙M

[0086] Where X represents the original feature matrix, X′ represents the feature matrix after the masking operation, and ⊙ represents the element-by-element multiplication of row-wise broadcasting. To protect the key semantic information, the feature of the source post node v0 is forced to be retained, that is, M[0] = 1; the enhanced view generated in this step is recorded as

[0087] Step 5.3: Enhance the view and Input graph encoder to generate graph-level representation h′ g and h″ g ; Use InfoNCE loss function to model semantic consistency:

[0088]

[0089] Among them, sim(·) is the cosine similarity function, τ is the temperature parameter, B is the batch size, Represents the graph-level features of the k-th sample in the current training batch;

[0090] Furthermore, the specific method of step 6 is:

[0091] Total loss function Defined as classification loss With contrastive learning loss The weighted sum of:

[0092]

[0093] The dynamic weight coefficient λ(t) adopts the cosine annealing scheduling mechanism and adaptively decays with the training cycle t∈[0,E]:

[0094]

[0095] Where η is the initial scaling factor and E is the total training cycle.

[0096] Compared with the prior art, the present invention has the following beneficial effects:

[0097] 1. Adopting a hierarchical graph coding architecture, the multi-scale extraction capability of propagation features is enhanced through the synergy of local attention and global topological perception.

[0098] 2. Adopting a dynamic gating mechanism, adaptively adjusting the feature fusion strategy, and optimizing the collaborative representation of heterogeneous modal features;

[0099] 3. Combined with a self-supervised contrastive learning method based on graph structure enhancement, it alleviates the impact of random perturbations in the propagation graph on detection results and improves the model's noise resistance.

[0100] 4. Adopt modular system design to support independent optimization and flexible replacement of each functional component, facilitating system function expansion according to actual needs. BRIEF DESCRIPTION OF THE DRAWINGS

[0101] Figure 1 Flowchart of a method according to an embodiment of the present invention.

[0102] Figure 2 Schematic diagram of the structure of an embodiment of the present invention. Specific implementation plan

[0103] In order to fully present the technical logic and implementation details of the present invention, the technical solution of the present invention will be systematically explained below in combination with the accompanying drawings and embodiments. Figure 1 The method flow chart of an embodiment of the present invention is shown, which specifically includes the following steps:

[0104] Step 1: Text feature extraction

[0105] To address the non-standard spelling and redundant noise issues in social media text, this step performs structured cleaning on the original text through a multi-stage processing flow: First, a regular expression matching mechanism is used to remove user identifiers containing the "@" symbol to eliminate the interference of usernames on semantic parsing; next, consecutive spaces and trailing punctuation marks are cleaned to retain valid semantic characters; finally, text normalization is performed, uniformly converting to lowercase and removing non-ASCII characters to form standardized input text.

[0106] After noise cleaning, the pre-processed text is fed into the pre-trained BERT model for contextual semantic encoding. Specifically, a classification marker [CLS] and a separation marker [SEP] are inserted at the beginning and end of the text to construct a structured sequence. The sequence length is then standardized to a fixed size m through dynamic truncation or zero padding. After this sequence is fed into BERT's multi-layer Transformer architecture, the multi-head self-attention mechanism is used to calculate the contextual association weights between tokens and generate a text semantic feature matrix. where d t =768 is the hidden layer dimension.

[0107] In order to further adapt to the requirements of false information detection tasks, the first row of feature vectors corresponding to the [CLS] tag is extracted As a global semantic representation. By constructing a nonlinear projection layer with trainable parameters, high-dimensional features are mapped to low-dimensional space: using the weight matrix W t and the bias term b t Perform linear transformation and apply ReLU activation function to finally generate a text feature vector with dimension d=64 This operation preserves the semantic representation ability of the pre-trained model while reconstructing discriminative features for the target task through learnable parameters.

[0108] Step 2: Social graph construction

[0109] The construction of social communication graph is based on source posts and hierarchical modeling to form a structured representation containing semantic features and topological relationships. First, based on the temporal communication chain of user interaction, a network of n i Node graph instance The root node corresponds to the original post, and the remaining n i -1 node represents the reply post. Node feature matrix It is composed of the TF-IDF feature vectors of all posts. This feature uses the word frequency-inverse document frequency statistical method to extract the text keyword distribution pattern and construct a discriminative semantic space with domain adaptability, where d g is the feature dimension.

[0110] To fully model the bidirectional impact of information dissemination, the edge set ε i Supports three construction modes: 1) Top-down (TD) mode through directed edges p→c Encode the direct reply relationship from parent node to child node; 2) Bottom-Up (BU) mode reversely defines the directed edge e c→p , highlighting the response of child nodes to parent nodes; 3) Undirected (UD) mode uses bidirectional edges Represents the interaction between nodes. Its implementation is due to the compatibility restrictions of the PyTorch Geometric framework on undirected edge identifiers. By adding reverse edges to simulate undirected connections, the number of edges is twice that of the TD / BU mode. The edge sets of the three modes can be formally defined as:

[0111]

[0112] The calculation of node degree is dynamically adapted according to the edge mode: in TD mode, the out-degree of each node is counted, i.e., deg(v j )=∑ k A i [j,k]; in BU mode, calculate the in-degree, i.e. deg(v j )=∑ k A i [k, j]; Although the UD mode expands the edge set, it still maintains the degree value consistency of the original propagation chain, that is, the node degree is calculated based on the initial edge definition, ensuring that the degree size is consistent in the three modes.

[0113] This construction method forms a multi-dimensionally encoded graph data object by integrating text features, communication topology and node influence, providing a joint representation space for graph neural networks that combines semantic associations and social interactions. Subsequent models can be constructed through the adjacency matrix A i With the feature matrix X i Collaborative computing enables interpretable modeling of propagation patterns.

[0114] Step 3: Graph structure feature extraction

[0115] After completing the construction of the social graph, this step uses the hierarchical graph neural network architecture to achieve in-depth extraction of the propagation structure features. First, through linear projection and layer normalization operations, the high-dimensional sparse TF-IDF features are mapped to the low-dimensional latent space. The calculation process is:

[0116] H (0) =ReLU(LayerNorm(XW p ))

[0117] The projection matrix Complete feature dimensionality reduction, layer normalization operation to improve the stability of model training, d0 is the dimension of the latent space. On this basis, the GAT-GCN cascade structure is used for feature enhancement: first, the local interaction pattern is captured by GAT, and the dynamic correlation weight between nodes is calculated using the multi-head attention mechanism. Assume that the number of attention heads is k, and the node v i The aggregated features are expressed as:

[0118]

[0119] In the formula Represents node v i The set of neighbor nodes of is the parameter matrix of the m-th head, is the attention coefficient. Then the nonlinear transformation is introduced through the ELU activation function to generate the intermediate feature matrix H (1) =ELU(H (1) ).

[0120] In order to capture the global topological characteristics, the GAT output features are further input into the GCN layer and a residual connection is established with the initial projection features:

[0121]

[0122] in is the normalized adjacency matrix, and They are the graph convolution parameter matrix and the skip connection parameter matrix respectively. This residual structure realizes the complementary transmission of multi-granularity information through the linear superposition of shallow and deep features.

[0123] Different from the conventional global pooling method, this step uses the attention gating mechanism to dynamically learn the node importance weights and suppress the interference of noise nodes on the graph-level representation. Specifically, the weight of each node is calculated by the scoring function g(·) composed of two fully connected layers. Then generate weighted graph embedding:

[0124]

[0125] Finally, feature space alignment is completed through nonlinear transformation:

[0126]

[0127] In the formula and is a learnable parameter matrix, ensuring that the output dimension d is consistent with the text feature h t This architecture significantly improves the representation capability of key nodes in the propagation path through the coordinated optimization of local attention perception and global topology modeling.

[0128] Step 4: Feature fusion and classification decision

[0129] In order to achieve the collaborative modeling of text semantics and graph structure features, this step realizes the adaptive fusion of heterogeneous features through a dynamic weight allocation mechanism. and graph features The learnable gating parameter θ is used to generate the normalized weight coefficients:

[0130] α=Sigmoid(θ)

[0131] This coefficient is constrained to the (0,1) interval through the Sigmoid function, dynamically adjusting the contribution ratio of the two types of features. On this basis, the original features are weighted and scaled:

[0132] h t′ =αh t ,h g′ =(1-α)h g

[0133] The weighted features are then concatenated along the dimension to construct a joint representation vector:

[0134]

[0135] This fusion strategy effectively balances the correlation between local semantic details and global propagation patterns through adaptive learning of parameter α, avoiding the problem of feature contribution solidification caused by traditional static weighting.

[0136] In the classification decision stage, the fusion features are input into the multi-layer perceptron to realize the category probability mapping. First, the high-order nonlinear features are extracted through the fully connected layer and the ReLU activation function:

[0137]

[0138] The false information detection task in the subsequent steps contains four label categories, so the four-category probability distribution is finally calculated through the softmax function:

[0139]

[0140] Model training uses the cross entropy loss function as the classification loss for end-to-end optimization, and the calculation formula is:

[0141]

[0142] Step 5: Graph Enhanced Contrastive Learning

[0143] To address the dynamic evolution of misinformation dissemination on social media and the scarcity of annotated data, this step proposes a degree centrality-guided hierarchical contrastive learning framework, which constructs semantic invariance constraints through a dual pathway of structural and feature enhancement. This approach, comprised of three core modules: topological robustness enhancement, feature integrity enhancement, and multi-view contrast optimization, aims to uncover essential patterns in social communication graphs and mitigate the risk of model overfitting.

[0144] In the topology robustness enhancement module, a centrality-weighted edge perturbation mechanism is adopted. First, define the node v j The logarithmic degree centrality of to balance scale differences:

[0145] C(v j )=log(deg(v j )+1)

[0146] Next, edge e is calculated based on node centrality i→j The enhanced weight of:

[0147]

[0148] The edge deletion strategy calculates dynamic probabilities based on weights and prioritizes removing edges with low centrality:

[0149]

[0150] where β d is the baseline deletion probability, s max With s mean are the maximum and mean values ​​of the edge weights of the entire graph, respectively. After generating the retained edge set ε′ through Bernoulli sampling, the edge addition strategy introduces random perturbations: candFilter unconnected edges in proportion β a Randomly added to ε′ to form an enhanced edge set ε″. This strategy generates an enhanced view by retaining the key propagation path and expanding the diversity path

[0151] In the feature integrity enhancement module, a centrality-weighted feature shielding mechanism is designed. j The shielding probability of is dynamically adjusted by the logarithmic degree centrality:

[0152]

[0153] where β m is the baseline masking probability. Then, a mask vector M is generated based on the masking probability, and the masking operation is performed on the feature matrix:

[0154] X′=X⊙M

[0155] In order to protect key semantic information, the features of the source post node v0 (i.e., M[0] = 1) are retained to avoid the negative impact of information loss on graph structure modeling. This strategy dynamically masks low-influence node features, forcing the model to capture discriminative patterns under partial information loss conditions and generate enhanced views.

[0156] In the multi-view comparison optimization module, the enhanced view and Input graph encoder to generate graph-level representation h′ g and h″ g . InfoNCE loss function is used to model semantic consistency:

[0157]

[0158] Where sim(·) is the cosine similarity function, τ is the temperature parameter, and B is the batch size. This loss function maximizes the similarity of positive sample pairs and suppresses the correlation of negative sample pairs, driving the encoder to learn enhanced invariant graph structure representation.

[0159] Step 6: Model training and testing

[0160] In order to collaboratively optimize the classification task and the self-supervised contrastive learning task, this step proposes a dynamic weighted loss balancing strategy. The overall objective function is defined as the classification loss With contrastive learning loss The weighted sum of:

[0161]

[0162] The dynamic weight coefficient λ(t) adopts the cosine annealing scheduling mechanism and adaptively decays with the training cycle t∈[0,E]:

[0163]

[0164] Where η is the initial scaling factor and E is the total training cycle. This design ensures that λ(t) satisfies the following decay characteristics: 1) In the initial stage, high-weighted contrastive learning constraints are used to enhance the robustness of graph structure representation; 2) In the final stage, the classification loss is fully dominant, preventing self-supervisory signals from interfering with classifier optimization.

[0165] By periodically adjusting the contrastive learning intensity, the model learns to enhance the invariant propagation pattern features in the early stage of training, and gradually focuses on refining the classification decision boundary in the later stage to achieve task-driven representation transfer.

[0166] To validate the effectiveness of this method, this step conducts performance tests on two social media misinformation detection benchmark datasets, Twitter15 and Twitter16. These datasets were constructed by Ma et al. based on the Twitter platform and employ a four-category annotation framework (verified misinformation, disproven misinformation, unverified misinformation, and non-misinformation) to capture the dynamic semantic evolution of misinformation during its spread.

[0167] Following existing work in this field, we used a stratified random sampling method to split the training and test sets into an 8:2 ratio, ensuring that the distribution of the four classes of samples was consistent with the original data. The statistical information of the divided dataset is shown in Table 1.

[0168] Table 1 Statistics of the dataset

[0169]

[0170] The experiment uses two evaluation metrics: global accuracy (Acc) and class-wise F1 scores (Class-wise F1). Acc measures the model's overall predictive accuracy and is calculated as the ratio of the sum of the number of true positive samples in the four categories to the total number of samples in the test set. Class-wise F1 independently calculates the harmonic mean of precision (Precision) and recall (Recall) for the four categories of TR, FR, UR, and NR to reveal differences in the model's recognition of different types of false information. The experimental results will be reported simultaneously with Acc and the four-category F1 scores, balancing the needs of global performance evaluation and fine-grained category analysis.

[0171] The specific statistical results are shown in Table 2, where bold indicates the best results and underlined indicates the suboptimal results. In terms of overall accuracy, the proposed DGCF method achieved the best performance on both datasets, reaching 86.3% and 90.4%, respectively, outperforming all baseline methods and improving by 1.1 percentage points over the suboptimal methods. This demonstrates the effectiveness of DGCF in multi-category misinformation detection tasks.

[0172] Table 2 Performance comparison of DGCF and baseline methods

[0173]

[0174]

[0175] To systematically verify the synergy of the core modules in the DGCF algorithm, this step performs ablation experiments on two datasets. By horizontally comparing the performance differences between the complete model and four control models, the contribution of each technical component is quantified. The control model is defined as follows:

[0176] (1) BaseGCN: A baseline architecture based on a two-layer GCN, which only utilizes the propagation graph structure features;

[0177] (2) DGCF-Graph: removes the text feature branch and retains the GAT-GCN concatenated graph encoder;

[0178] (3) DGCF-GCL: integrates the degree centrality comparative learning module based on DGCF-Graph;

[0179] (4) DGCF-Concat: replaces dynamic gated fusion with direct feature concatenation.

[0180] The experimental results are shown in Table 3. The results show that: 1) Compared with the basic GCN encoder, the hierarchical graph encoder effectively improves the representation ability of social communication structure through the collaborative modeling of GAT and GCN; 2) The overall accuracy of DGCF-GCL on both datasets is better than that of DGCF-Graph, verifying the optimization effect of contrastive learning on graph representation learning; 3) The dynamic gating mechanism effectively alleviates the negative effects caused by the dominance of single modal information in the static splicing strategy through adaptive adjustment of cross-modal weights; 4) The complete DGCF method achieves comprehensive optimal performance in cross-dataset and cross-category detection, verifying the collaborative optimization effect of multiple components in the technical solution.

[0181] Table 3 Ablation study of DGCF on two datasets

[0182]

[0183]

[0184] The hardware and software configuration information of the experiment is shown in Table 5.

[0185] Table 5 Experimental hardware and software configuration information

[0186]

[0187] In addition to the implementation details mentioned in the above step description, the following settings are also adopted in the experiment:

[0188] In step 1, the standardized length of the text sequence is 50 tokens; the pre-trained model selected is BERT-base-uncased; the text feature extraction module contains two fully connected layers, with the input dimension being the pre-trained model hidden layer dimension of 768 and the intermediate layer dimension of 2048, and finally projected into a 64-dimensional semantic space; a random dropout rate of 0.2 is set during network training to improve generalization ability.

[0189] In step 2, node features are encoded into 5000-dimensional sparse vectors using the TF-IDF method; the edge set uses the undirected mode (UD) to model the bidirectional propagation effect.

[0190] In step 3, the feature preprocessing stage compresses the 5000-dimensional input to 1024 dimensions through a linear projection layer; the GAT layer is configured with 4 independent attention heads, each of which outputs 256-dimensional local interaction features, which are spliced ​​into a 1024-dimensional intermediate representation after ELU activation; the GCN layer further aggregates global topological features to 512 dimensions; the gated network of the attention pooling layer contains two linear layers (512→256→1); the pooled 512-dimensional graph-level embedding is output by a feature compression module (linear layer 512→256→64, with a probability of 0.4 random inactivation) to output a 64-dimensional graph structure feature vector.

[0191] In step 4, the gating parameter is initialized to 0.5.

[0192] In step 5, the baseline structure perturbation probability is uniformly set to 0.2, and the comparative learning temperature coefficient is 0.07.

[0193] In step 6, the initial scaling factor of the contrastive loss weight is set to 0.1 and 0.2 on the Twitter15 and Twitter16 datasets, respectively. The Adam optimizer is used for parameter update, with an initial learning rate of 0.004 and a weight decay coefficient of 0.001. The batch size is fixed to 64, and early stopping is implemented for training within 50 training cycles. The global random seed is fixed to 2025.

[0194] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should be regarded as within the scope of protection of the present invention.

Claims

1. A method for detecting false information based on text and social graphs, the method comprising the following steps: Step 1: Text feature extraction; Perform multi-level cleaning preprocessing on raw social media text to eliminate interference from user identifiers and non-standard characters and construct a standardized input sequence. A pre-trained language model is used for contextual semantic encoding, and a multi-layer Transformer architecture is used to extract global semantic representations. A nonlinear projection layer is designed to reduce the dimensionality of high-dimensional features, generating discriminative text feature vectors suitable for the task of false information detection. Step 2: Build a social graph; A hierarchical communication graph structure is constructed based on source posts, with nodes representing user interaction relationships. The feature matrix uses statistical methods to extract the distribution of text keywords. A multimodal edge set is defined to model the influence of two-way communication, supporting three construction strategies: top-down, bottom-up, and undirected connections. Based on the dynamic fitness centrality calculation of edge patterns, a graph data structure that combines semantic associations and social interactions is formed, providing a joint representation space for subsequent graph neural networks. Step 3: Graph structure feature extraction; A hierarchical graph neural network architecture is constructed, achieving dual perception of local interaction patterns and global topological features through the sequential concatenation of GAT and GCN. A multi-head attention mechanism is used to dynamically calculate the association weights between nodes to capture key local information in the propagation path. Graph convolutional layers are used to establish residual connections, fusing shallow and deep features to enhance global topological modeling capabilities. An attention gating mechanism is introduced to evaluate node importance, and a robust graph embedding representation is generated through weighted aggregation. Finally, feature space dimension alignment is achieved through nonlinear transformation. Step 4: Feature fusion and classification decision; A dynamic gating mechanism is designed to achieve adaptive fusion of text and image features, dynamically adjusting the weight ratio of heterogeneous features through learnable parameters. A dimension splicing strategy is used to construct a joint representation vector, balancing the relevance of semantic details and propagation patterns. The fused features are mapped to the classification space through multiple layers of nonlinear transformations, and end-to-end false information detection decisions are achieved through probability distribution generation. Step 5: Image enhancement contrastive learning; A node degree-guided self-supervised contrastive learning framework is constructed, which generates an enhanced view of the graph structure features obtained in step 3 through a dual path of topological perturbation and feature masking. Dynamically adjust edge deletion and feature masking probabilities based on node influence, retain key propagation paths, and enhance the model's adaptability to partial observation data. Design a multi-view comparison optimization strategy that maximizes semantic consistency constraints to drive the graph encoder to learn enhanced invariant feature representations, and combine supervision signals to achieve a synergistic improvement in classification performance and generalization ability; Step 6: Model training and testing; A dynamic weighted loss balancing strategy is designed, using a cosine annealing scheduling mechanism to coordinate the optimization strengths of the classification task and self-supervised contrastive learning. Experiments are conducted on a mainstream social media misinformation detection benchmark dataset, using both global accuracy and per-class F1 scores for performance evaluation. Finally, ablation experiments systematically validate the synergistic optimization effect of the hierarchical graph coding architecture, dynamic gating fusion, and contrastive learning module.

2. The method for detecting false information based on text and social graph according to claim 1, wherein: The specific method of step 1 is: Step 1.1: Use a regular expression matching mechanism to remove user identifiers containing the "@" symbol to eliminate the interference of usernames in semantic parsing; Step 1.2: Clean up consecutive spaces and trailing punctuation, retaining valid semantic characters, and perform text normalization, converting them to lowercase and removing non-ASCII characters to form the normalized input text. ASCII characters refer to English and common symbols, excluding special characters such as Chinese characters and emoticons. Step 1.3: Input the pre-processed text into the pre-trained BERT model for contextual semantic encoding; A classification marker [CLS] and a separation marker [SEP] are inserted at the beginning and end of the text to construct a structured sequence. The sequence length is standardized to a fixed size m through dynamic truncation or zero padding operations. After the sequence is input into the multi-layer Transformer architecture of BERT, the multi-head self-attention mechanism is used to calculate the contextual association weights between word units to generate a text semantic feature matrix. where d t is the mark symbol for the hidden layer dimension t to represent the text-related features; Step 1.4: Extract the first row of feature vectors corresponding to the [CLS] tag As a global semantic representation; by constructing a nonlinear projection layer with trainable parameters, high-dimensional features are mapped to low-dimensional space: using the weight matrix W t and the bias term b t Perform linear transformation and apply ReLU activation function to finally generate text feature vector 3. The method for detecting false information based on text and social graph according to claim 1, wherein: The specific method of step 2 is: Step 2.1: Construct a temporal propagation chain based on user interaction containing n i Node graph instance in Represents a set of nodes, including the root node corresponding to the original post and the remaining n nodes corresponding to the reply posts i -1 node; ε i Represents the edge set. Node feature matrix It is composed of the TF-IDF feature vectors of all posts. This feature uses the word frequency-inverse document frequency statistical method to extract the text keyword distribution pattern and construct a discriminative semantic space with domain adaptability, where d g is the feature dimension; Step 2.2: Construct edge set ε i , including three build modes: 1) Top-down mode: through directed edges e p→c Encode parent node v p To child node v c Direct reply relationship; 2) Bottom-up mode: define the directed edge e in reverse c→p , highlight the child node v c For parent node v p Response effect; 3) Undirected mode: using bidirectional edges Represents the interaction between nodes; The edge sets of the three modes can be formally defined as: Node degree deg(v j ) is dynamically adapted according to the edge mode: in the top-down mode, the out-degree of each node is counted, i.e., deg(v j )=∑ k A i [j,k]; calculate the in-degree in the bottom-up mode, that is, deg(v j )=∑ k A i [k,j]. Among them, A i for The adjacency matrix, A i [j,k]=1 indicates that there is an edge from node v j Points to node v k Although the undirected mode expands the edge set, it still maintains the degree value consistency of the original propagation chain, that is, the node degree is calculated based on the initial edge definition, ensuring that the degree size is consistent in the three modes.

4. The method for detecting false information based on text and social graph according to claim 1, wherein: The specific method of step 3 is: Step 3.1: Node feature matrix of social communication graph First, through linear projection and layer normalization operations, the high-dimensional sparse TF-IDF features are mapped to a low-dimensional latent space, and then nonlinearly transformed through the ReLU activation function. The calculation process is as follows: H (0) =ReLU(LayerNorm(XW p )) Among them, H (0) Represents the feature matrix after dimensionality reduction; LayerNorm() represents the layer normalization operation; is the projection matrix, d0 is the latent space dimension; Step 3.2: Use GAT-GCN tandem for feature enhancement: First, use GAT to capture local interaction patterns and use the multi-head attention mechanism to calculate the dynamic association weights between nodes; let the number of attention heads be k, and the node v i Aggregation features Expressed as: In the formula, || represents the symbol || represents the splicing operation on the feature dimension; Represents node v i The set of neighbor nodes of Represents the neighbor node v in the mth head j For node v i The attention weight of Represents node v j In the feature matrix H (0) The corresponding eigenvector in ; is the parameter matrix of the mth head; then the nonlinear transformation is introduced through the ELU activation function to generate the intermediate feature matrix H (1) =ELU(H (1) ); Step 3.3: Input the GAT output features into the GCN layer and establish a residual connection with the initial projection features: Among them, H (2) represents the intermediate feature matrix obtained in this step, is the normalized adjacency matrix, and They are the graph convolution parameter matrix and the skip connection parameter matrix respectively. This residual structure realizes the complementary transmission of multi-granularity information through the linear superposition of shallow and deep features; Step 3.4: Use the attention gating mechanism to dynamically learn the node importance weights, and calculate the weight of each node through the Sigmoid function and the scoring function g(·) composed of two fully connected layers Then generate weighted graph embedding in, Represents node v i In the feature matrix H (2) The corresponding eigenvector in ; Finally, feature space alignment is completed through nonlinear transformation: Where, and is a learnable parameter matrix, ensuring that the output dimension d is consistent with the text feature h t Stay consistent.

5. The method for detecting false information based on text and social graph according to claim 1, wherein: The specific method of step 4 is: Step 4.1: Targeting text features and graph features The learnable gating parameter θ is used to generate the normalized weight coefficient α: α=Sigmoid(θ) This coefficient is constrained to the (0,1) interval through the Sigmoid function, dynamically adjusting the contribution ratio of the two types of features; on this basis, the original features are weighted and scaled: h t′ =ah t ,h g′ =(1-a)h g Among them, h t′ and h g′ They are weighted and scaled text features and graph features respectively; Step 4.2: The weighted features are then concatenated along the dimension to construct a joint representation vector h fuse : Step 4.3: Classification decision; Extract high-order nonlinear features z through the fully connected layer and ReLU activation function: in, and They are the weight matrix and bias term when performing linear transformation; The false information detection task includes four label categories, and the four-category probability distribution is finally calculated using the softmax function: in, is the predicted label, and They are the weight matrix and bias term when performing linear transformation; Model training uses the cross entropy loss function as the classification loss for end-to-end optimization, and the calculation formula is:

6. The method for detecting false information based on text and social graph according to claim 1, wherein: The specific method of step 5 is: Step 5.1: Based on node v j The degree of deg(v j ), define its logarithmic degree centrality C(v j ) to balance the scale differences: C(v j )=log(deg(v j )+1) Step 5.2: Calculate edge e based on node centrality i→j The enhanced weight s(v i ,v j ): In the edge deletion strategy, edge e is calculated based on the enhanced weight i→j The deletion probability p drop (e i→j ), this strategy prioritizes removing low-weight edges: Among them, β d is the baseline deletion probability; s max With s mean are the maximum and mean values ​​of the edge weights of the entire graph, respectively; ∈ is a smoothing factor to avoid the denominator being zero; after generating the retained edge set ε′ through Bernoulli sampling, the edge addition strategy introduces random perturbations to generate enhanced views The random perturbation method is: from the candidate edge set ε cand Filter unconnected edges in proportion β a Randomly added to ε′ to form an enhanced edge set ε″; Step 5.3: Centrality-weighted feature masking; Node v j The shielding probability p mask (v j ) is dynamically adjusted by log degree centrality: Among them, β m is the baseline shielding probability; C max with C mean are the maximum and mean values ​​of the logarithmic degree centrality of the nodes in the entire graph, respectively. Then, a mask vector M is generated based on the masking probability, and a masking operation is performed on the feature matrix X: X′=X⊙M Where X represents the original feature matrix, X′ represents the feature matrix after the masking operation, and ⊙ represents the element-by-element multiplication of row-wise broadcasting. To protect the key semantic information, the feature of the source post node v0 is forced to be retained, that is, M[0] = 1; the enhanced view generated in this step is recorded as Step 5.3: Enhance the view and Input graph encoder to generate graph-level representation h′ g and h′ g '; InfoNCE loss function is used to model semantic consistency: Among them, sim(·) is the cosine similarity function, τ is the temperature parameter, B is the batch size, Represents the graph-level features of the k-th sample in the current training batch.

7. The method for detecting false information based on text and social graph according to claim 1, wherein: The specific method of step 6 is: Total loss function Defined as classification loss With contrastive learning loss The weighted sum of: The dynamic weight coefficient λ(t) adopts the cosine annealing scheduling mechanism and adaptively decays with the training cycle t∈[0,E]: Where η is the initial scaling factor and E is the total training cycle.

Images