A cross-domain rumor detection method based on field decoupling

Abstract

The application belongs to the field of new generation information technology, and particularly relates to a cross-domain rumor detection method based on field decoupling, comprising the following steps: S1, constructing an initial propagation tree for source tweets and associated user comments; reconstructing a denoising propagation tree; S2, modeling a propagation mode to obtain a propagation mode representation; extracting semantic features of the source tweets to obtain semantic representations of the source tweets; S3, inputting the propagation mode representation into a fact decoupler to decouple fact binding features and false stance noise features; S4, inputting the semantic representations of the source tweets into a field decoupler to decouple field invariant features and field specific features; S5, respectively reweighting the fact binding features and the field specific features, and then performing dynamic weighted fusion to generate a clue dominant representation; and S6, splicing the clue dominant representation and the field invariant features, and then inputting the spliced result into a false information detector for classification to output a rumor detection result. The method can realize high-precision cross-domain rumor detection.

Description

Technical Field

[0001] This invention belongs to the field of next-generation information technology, and in particular relates to a cross-domain rumor detection method based on domain decoupling. Background Technology

[0002] With the rapid development of social media platforms, the speed and reach of information dissemination have reached unprecedented levels. However, the rapid and widespread spread of rumors on social media poses an increasing threat to people's daily lives, public trust, and even social order. To effectively curb this risk, the development of automated rumor detection technologies is imperative.

[0003] In existing technologies, rumor detection primarily relies on two types of data: textual features based solely on news content, and social context data including user activity. Based on this data, traditional in-domain rumor detection techniques determine the authenticity of information by extracting effective features from news articles or user comments. However, when facing emerging topics or unknown domains, due to significant differences in data distribution, directly deploying models trained in the source domain to the target domain often leads to a substantial performance drop.

[0004] To address this challenge, researchers have focused on cross-domain knowledge transfer techniques, primarily concentrating on domain-invariant feature extraction. Existing techniques mainly include domain adversarial training methods and class-aware methods. The former utilizes techniques such as adversarial learning and maximum mean difference to align feature distributions across different domains; the latter constructs source domain data consistent with the target distribution through techniques such as pseudo-labels and data selection. However, these existing techniques have significant limitations: they heavily rely on the participation of target domain data, leading to failure in scenarios where the target domain is completely unknown, and the additional requirement for target domain samples significantly increases training and deployment costs.

[0005] Although generalized rumor detection has become a research hotspot, existing technologies still face three major challenges in practical applications. These problems are complex in origin and difficult to solve:

[0006] 1. The problem of interference in responses. Existing detection methods based on social context typically rely on the user group's stance to infer authenticity. However, actual propagation trees are filled with a large number of stance-neutral, noisy comments. These responses (such as...) Figure 1 The responses (R2, R3, R5, and R6) within the red border area have low relevance to the core topic of the source tweet. They not only fail to provide effective verification clues but also distort the topological features of the propagation structure. Due to the diversity of natural language, accurately distinguishing between "position-related" and "position-independent" is highly semantically ambiguous. Without effective denoising, the model is prone to capturing incorrect propagation patterns, leading to a decrease in detection accuracy.

[0007] 2. Domain Heterogeneity Issues. Domain heterogeneity stems not only from differences in dissemination patterns across different domains, but also from the mismatch between the dissemination pattern and the semantic features of the source tweet. For example, in some domains, although user replies (such as...) Figure 1 In the rumor propagation structure shown, R1, R4, R7, R 12 The responses (such as those from the original tweet R0) expressed opinions on the topic, but the semantics of the discussion may have deviated. This inconsistency between the propagation structure and semantic content (i.e., domain heterogeneity) makes it difficult for a single feature extractor to capture robust cross-domain features simultaneously. How to alleviate this misalignment between structure and semantics during the modeling process is a key challenge in improving the model's generalization ability.

[0008] 3. Knowledge Dependency Issues. The degree of knowledge dependence varies significantly across different domains. For example, in the news domain, the dissemination pattern is highly semantically related to the source message, and model decisions heavily rely on semantic coherence. In the entertainment domain, however, dissemination is often accompanied by emotional expression, and the dissemination pattern is relatively separate from the semantics of the source message. This dynamic change in knowledge dependence makes it difficult for general-purpose models to maintain stable discriminative capabilities across different scenarios. Existing technologies typically employ fixed feature fusion strategies, which cannot adapt to this cross-domain dependence variation, thus limiting the model's adaptability.

[0009] Therefore, how to effectively decouple fact-binding features from domain-specific features while removing interference from comments irrelevant to one's stance, and achieve high-precision cross-domain rumor detection, has become an urgent problem to be solved. Summary of the Invention

[0010] To address the aforementioned shortcomings of existing technologies, the present invention aims to provide a cross-domain rumor detection method based on domain decoupling, which can effectively decouple fact-binding features from domain-specific features while removing interference from position-irrelevant comments, thereby achieving high-precision cross-domain rumor detection.

[0011] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0012] A cross-domain rumor detection method based on domain decoupling includes the following steps:

[0013] S1. Construct an initial propagation tree for the source tweet and its associated user comments; use a large language model to classify the stance of each user comment in the initial propagation tree, with stance categories including neutral and non-neutral; delete comment nodes with a neutral stance category, and for the deleted non-leaf neutral nodes, reconnect all their child nodes to their parent nodes to reconstruct the denoised propagation tree.

[0014] S2. Extract the adjacency matrix of the denoised propagation tree obtained in S1 as a structural prior and inject it into the attention calculation of the graph Transformer model to model the propagation pattern and obtain the propagation pattern representation; and use the pre-trained language model to extract the semantic features of the source tweet to obtain the semantic representation of the source tweet.

[0015] S3. Input the propagation pattern representation obtained in S2 into the trained fact decoupler to decouple the fact binding features and false position noise features.

[0016] S4. Input the semantic representation of the source tweet obtained in S2 into the trained domain decoupler to decouple the domain-invariant features and domain-specific features.

[0017] S5. After reweighting the fact binding features obtained in S3 and the domain-specific features obtained in S4, dynamic weighted fusion is performed to generate a clue-dominated representation.

[0018] S6. After concatenating the dominant cue representation generated in S5 with the domain-invariant features obtained in S4, input the concatenation into the trained fake information detector for classification and output the rumor detection results.

[0019] S3 and S4 can be interchanged.

[0020] Compared with the prior art, the present invention has the following advantages:

[0021] 1. A stance-aware propagation tree denoising method significantly improves structural modeling quality. By introducing a large-scale language model to perform fine-grained stance classification of user comments and selectively removing neutral comment nodes, this approach effectively eliminates a large amount of noise interference in the propagation structure. Compared with existing techniques that directly use the original propagation tree, this method not only simplifies the complexity of the graph structure but, more importantly, ensures that subsequent propagation pattern modeling is based solely on stance-related comments with discriminative value, thereby significantly improving the purity and discriminative power of structural features.

[0022] 2. Deep fusion of graph structure and semantic features enhances multimodal representation capabilities. This approach innovatively injects the adjacency matrix of the denoised propagation tree as a structural prior into the attention calculation of the graph Transformer, while simultaneously utilizing a pre-trained language model to extract deep semantic features from the source tweet. Compared to the simple concatenation or independent processing of structural and semantic features in traditional methods, this fusion strategy enables deep interaction between propagation patterns and textual content, thereby obtaining richer and more comprehensive information representations.

[0023] 3. A dual decoupling mechanism effectively separates interfering factors from core discriminative features. Through the collaborative work of a fact decoupler and a domain decoupler, this scheme achieves a refined decomposition of propagation pattern representation and semantic representation. The fact decoupler decomposes the propagation pattern into fact-bound features and false position noise features, while the domain decoupler decomposes the semantic features into domain-invariant features and domain-specific features. This dual decoupling mechanism effectively overcomes the feature entanglement problem in existing technologies, enabling the model to focus on the most essential discriminative information and significantly improving cross-domain generalization ability.

[0024] 4. Dynamic weighted fusion strategy for cue-driven adaptive decision-making. The dynamic weighted fusion mechanism designed in this scheme adaptively adjusts the fusion weights of fact-binding features and domain-specific features based on the differences in the importance of various features across different domains and specific contexts. Compared to the fixed-weight fusion strategies in existing technologies, this method can more flexibly capture the most discriminative cues, thereby generating more accurate cue-driven representations and providing strong support for the final rumor detection decision.

[0025] 5. Synergistic utilization of domain-invariant features and cue-dominant representations enhances detection robustness. By concatenating cue-dominant representations with domain-invariant features before inputting them into the detector, this scheme achieves an organic combination of domain-specific knowledge and cross-domain general knowledge. This design retains the discriminative ability specific to a particular domain while ensuring the model's stability when facing unknown domains, effectively addressing the problem of insufficient domain adaptability in existing technologies.

[0026] In summary, this method can effectively decouple fact-binding features from domain-specific features while removing interference from position-irrelevant comments, thus achieving high-precision cross-domain rumor detection.

[0027] Preferably, in S1, the non-neutral stance categories include support, opposition, and questioning; the step of classifying the stance of each user comment in the initial propagation tree using a large language model includes: calling the large language model interface, taking the source tweet and a single user comment as input, and using a prompt template to guide the large language model to classify the stance of each user comment into one of support, opposition, questioning, or neutral.

[0028] Preferably, in S1, reconnecting all child nodes to their parent nodes includes: traversing the initial propagation tree using a breadth-first search strategy; when the stance category of the currently traversed comment node is determined to be neutral, deleting the neutral node from the initial propagation tree and pointing the parent pointers of all child nodes of the neutral node to the parent node of the neutral node before deletion, until all neutral nodes have been processed to obtain the denoised propagation tree.

[0029] Compared to the traditional approach of simply deleting neutral nodes and discarding all their subtrees, this strategy preserves potentially valuable non-neutral comments downstream of neutral nodes through a parent pointer redirection mechanism, preventing the breakdown of crucial evidence chains downstream due to the neutrality of upstream nodes. This structurally preserving reconstruction ensures the semantic coherence and information integrity of the propagation path. Especially in the typical pattern of neutral quotations followed by subsequent rebuttals in rumor propagation, it effectively prevents the loss of important counter-evidence, thereby significantly improving the information fidelity and detection reliability of the denoised propagation tree.

[0030] Preferably, in S2, the propagation mode is modeled to obtain a propagation mode representation, including:

[0031] For a denoising propagation tree G=(V,E), where the set of nodes is... Including the source tweet and the sequence of user comments ;Will The text content of all nodes in the model is input into the BERT model to obtain the feature embeddings. ;

[0032] The features are embedded into the x-input graph Transformer model, and the self-attention weights of the h-th attention head are... for:

[0033] ;

[0034] in, Indicates the first The self-attention weights of the heads; A is the adjacency matrix of the denoising propagation tree; , The first The query matrix and key matrix of each attention head; Dimensions for each attention head;

[0035] By combining the self-attention output with the residual connection, and then passing it through a feedforward network and pooling operations, a propagation mode representation is obtained. :

[0036] ;

[0037] in, Indicates pooling; Indicates a feedforward network; For the number of heads; This is the value matrix of the h-th attention head; This is the learnable parameter matrix.

[0038] This approach, by directly embedding the graph structure prior (adjacency matrix A) into the attention mechanism's computation process, rather than simply fusing them at the input or output layers, achieves fine-grained coupling of structural and semantic information at the attention level. This allows the model to not only rely on textual semantic similarity when calculating inter-node correlations but also to enforce respect for real-world propagation topologies, such as the temporal order and dependencies between parent and child comments. This effectively alleviates the problems of pure text models ignoring propagation paths and being susceptible to semantic interference. It also outperforms existing methods that only use graph convolutional networks to process structure but lack deep semantic interaction, significantly improving the ability to model dynamic patterns of rumor propagation.

[0039] Preferably, the trained fact decoupler is obtained through the following training process:

[0040] S3.1 Construct a fact decoupler, the fact decoupler including a first gated branch for extracting fact binding features. and the second gating branch used to extract false position noise features The parameters are respectively ;

[0041] S3.2 Input the propagation pattern representation in the training data into the fact decoupler, and obtain the hidden representation through the self-attention mechanism. ;

[0042] S3.3 In the first gating branch, a gating mechanism is used to extract information from the hidden representation. Extract fact binding features from the data, including the gating vector. The formula for calculation is:

[0043] ;

[0044] in, It is the sigmoid function; , , , These are learnable parameters; To correct the activation function of the linear unit; where the superscript d is the hidden layer dimension, and the superscript dp is the projection dimension of the query, key, and value vectors in the self-attention mechanism;

[0045] The gate vector With the hidden representation Element-wise multiplication is performed, followed by layer normalization and multilayer perceptron mapping to obtain fact-binding features. :

[0046] ;

[0047] in, Representation layer normalization; It is a multilayer perceptron;

[0048] S3.4 In the second gating branch, the same gating structure as the first gating branch is adopted, using the gating vector calculated from another set of learnable parameters. Extracting false position noise features ;

[0049] S3.5 Construct the adversarial training objective function for the fact decoupler. :

[0050] ;

[0051] ;

[0052] ;

[0053] in, This indicates that the fact is bound to a feature. Input fraud detector The resulting classification loss; This indicates that the false position noise feature will be used. Input the fake information detector The resulting adversarial loss, where λ is a hyperparameter. , , These are the learnable parameters of the first gating branch, the second gating branch, and the fake information detector, respectively; x∈ y∈ , These represent the sample set and the label set, respectively. The cross-entropy loss function;

[0054] S3.6 Introduce reconstruction loss constraints to bind the facts to features. With the aforementioned false position noise features After concatenation, the hidden representation is reconstructed using a reconstruction function. Reconstruction loss for:

[0055] ;

[0056] Where [;] represents the vector concatenation operation;

[0057] S3.7, Combine the aforementioned adversarial training objective function and the reconstruction loss Total loss The fact decoupler is trained, and during the training process, an adversarial optimization between the second gating branch and the fake information detector is achieved through a gradient inversion layer until the model converges.

[0058] This approach, through a synergistic mechanism of adversarial decoupling and reconstruction constraints, achieves an explicit separation of factual content and stance noise in the representation of propagation patterns. It ensures that fact-bound features retain core factual clues sufficient to support rumor detection, such as event elements, time, and location, while simultaneously eliminating interference signals induced by false stances, such as emotional expressions and misleading rhetoric. This significantly improves the domain robustness and semantic purity of the features. Compared to traditional decoupling methods that rely solely on reconstruction loss or unsupervised clustering, this strategy guides the model to learn more discriminative and thoroughly decoupled factual representations through explicit adversarial objectives, laying a crucial foundation for subsequent cross-domain detection.

[0059] Preferably, the trained domain decoupler is obtained through the following training process:

[0060] S4.1 Construct a domain decoupler, which includes a third branch for extracting domain-invariant features and a fourth branch for extracting domain-specific features;

[0061] S4.2, Semantic representation of source tweets in training data The domain decoupler is input, and using the same gating mechanism, layer normalization, and multilayer perceptron mapping steps as in steps S3.3 to S3.4, domain-invariant features are obtained. Domain-specific features ;

[0062] S4.3 Construct a domain classifier to classify the domain-specific features. Input the domain classifier to predict its domain label and calculate the corresponding classification loss. Introducing reconstruction loss To constrain the domain-invariant features With the domain-specific features Capable of jointly reconstructing the semantic representation of the source tweet back to the input ;

[0063] S4.4, Combining Classification Loss and reconstruction loss Total loss The domain decoupler is trained, and during the training process, an adversarial optimization between the domain-specific feature branch and the domain classifier is achieved through a gradient inversion layer until convergence.

[0064] This setup, through the collaborative design of architecture reuse and domain adversarial training, effectively enhances the cross-domain robustness and transfer adaptability of propagation representation, providing key support for building a general rumor detection system.

[0065] Preferably, in S5, the fact-binding features and domain-specific features are reweighted, including:

[0066] The fact binding feature The reweighted mean is calculated using the first multilayer perceptron. and weighted variance And generate fact reweighted features according to the following formula. :

[0067] ;

[0068] in, This indicates element-wise multiplication.

[0069] And using the same calculation method, a domain-specific reweighted algorithm is generated. .

[0070] This setup, by introducing learnable feature-level dynamic reweighting, effectively alleviates the model sensitivity problem caused by excessive differences in the strength of factual features among different samples. For example, weak factual signals such as veiled statements are easily ignored, while strong noise signals such as high sentiment expressions easily dominate the decision. This mechanism adaptively adjusts the contribution of features based on their own distribution statistics, making the model more focused on relatively significant and stable factual components. Thus, while maintaining the decoupled structure, it further improves the discriminative stability and robustness of features, and is especially suitable for real rumor scenarios with diverse factual expressions and large signal-to-noise ratio fluctuations.

[0071] Preferably, in S5, the dynamic weighted fusion to generate a cue-dominant representation includes:

[0072] Reweighting the facts With the domain-specific reweighted features As input, the dynamic fusion weights are calculated using the softmax function after mapping through a third-layer perceptron. and :

[0073] ;

[0074] The features are weighted and fused according to the following formula to obtain the cue-dominated representation. :

[0075] ;

[0076] in, This indicates vector concatenation.

[0077] This setup, through a weighted fusion strategy based on dynamic decision-making of feature content, overcomes the limitations of fixed-ratio fusion or simple splicing. In different samples, the model can autonomously determine the relative importance of fact-binding information and domain-specific context to the current discrimination task, thereby achieving semantic-level adaptive integration of multi-source decoupled features and significantly improving the final representation's ability to focus on the essence of rumors and its adaptability to different scenarios.

[0078] Preferably, the fake information detector is a classifier used for binary classification tasks.

[0079] Preferably, the overall training objective function of the method is:

[0080] ;

[0081] in, The binary cross-entropy classification loss for the false information detector; The total loss of the fact decoupler; The total loss of the domain decoupler; is a hyperparameter used to balance the detection loss with the training objectives of the two types of decoupled modules. Attached Figure Description

[0082] To make the objectives, technical solutions, and advantages of the invention clearer, the invention will now be described in further detail with reference to the accompanying drawings, wherein:

[0083] Figure 1 This is an illustrative diagram for the background art;

[0084] Figure 2 This is a schematic diagram of the system framework when this method is implemented;

[0085] Figure 3 This is a flowchart of the method;

[0086] Figure 4 This is a schematic diagram of the ablation results in Example 2;

[0087] Figure 5 This is a T-SNE visualization result of the MONSTRSE method in Example 2;

[0088] Figure 6 This is a T-SNE visualization result of the UTBD method in Example 2;

[0089] Figure 7 This is a schematic diagram illustrating the influence of hyperparameters ω1, ω2, and ω3 on the UTBD method in Example 2. Detailed Implementation

[0090] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0091] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 a part of the embodiments of the present invention, not all of them. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0092] It should be noted that similar reference numerals and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the figures, or the orientation or positional relationship commonly used when the product is in use. They are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third," etc., are only used to distinguish descriptions and should not be construed as indicating or implying relative importance. In addition, the terms "horizontal," "vertical," etc., do not indicate that the component is required to be absolutely horizontal or suspended, but can be slightly tilted. For example, "horizontal" simply means that its direction is more horizontal than "vertical," and does not mean that the structure must be completely horizontal, but can be slightly tilted. In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0093] Example 1

[0094] This invention provides a cross-domain rumor detection method based on domain decoupling. For better understanding, the model architecture used in the specific implementation of this method is briefly described, such as... Figure 2As shown, the framework includes a position-independent response filtering stage and a domain generalization stage, the latter of which includes a Domain Heterogeneity Mitigation Module (DHMM) and a Domain Knowledge Fusion Module (DKFM), which correspond to the steps of this method, respectively. The specific correspondence will be explained in the following description.

[0095] like Figure 3 As shown, this method includes the following steps:

[0096] S1. Construct an initial propagation tree for the source tweet and its associated user comments; use a large language model to classify the stance of each user comment in the initial propagation tree, with stance categories including neutral and non-neutral; delete comment nodes with a neutral stance category, and for the deleted non-leaf neutral nodes, reconnect all their child nodes to their parent nodes to reconstruct the denoised propagation tree.

[0097] In specific implementation, the non-neutral stance categories include support, opposition, and questioning; the step of classifying the stance of each user comment in the initial propagation tree using a large language model includes: calling the large language model interface, taking the source tweet and a single user comment as input, and using a prompt template to guide the large language model to classify the stance of each user comment into one of support, opposition, questioning, or neutral.

[0098] The step of reconnecting all its child nodes to their parent nodes includes: traversing the initial propagation tree using a breadth-first search strategy; when it is determined that the stance category of the currently traversed comment node is neutral, deleting the neutral node from the initial propagation tree and pointing the parent pointers of all the child nodes of the neutral node to the parent node of the neutral node before deletion, until all neutral nodes have been processed, thus obtaining the denoised propagation tree.

[0099] Compared to the traditional approach of simply deleting neutral nodes and discarding all their subtrees, this strategy preserves potentially valuable non-neutral comments downstream of neutral nodes through a parent pointer redirection mechanism, preventing the breakdown of crucial evidence chains downstream due to the neutrality of upstream nodes. This structurally preserving reconstruction ensures the semantic coherence and information integrity of the propagation path. Especially in the typical pattern of neutral quotations followed by subsequent rebuttals in rumor propagation, it effectively prevents the loss of important counter-evidence, thereby significantly improving the information fidelity and detection reliability of the denoised propagation tree.

[0100] In social context-driven misinformation detection, the user group stance patterns presented by the propagation tree are widely used to assist in judging the authenticity of information. Typical stance categories include support, deny, question, and neutral. Among them, supportive, opposing, and questioning responses not only reflect the public's attitude towards the source post information, but also often implicitly contain judgments on its credibility and clues of doubt. In contrast, a large number of comments with a neutral stance usually manifest as emotional venting or off-topic discussions unrelated to the authenticity of the event. These responses lack substantial semantic connection with the authenticity of the source tweet and can therefore be regarded as "stance-irrelevant" noise nodes. Directly incorporating them into the propagation structure model can easily obscure effective stance clues, interfere with propagation pattern learning, and increase structural redundancy and time overhead in the model inference stage. Based on this, the core task of this stage can be described as: judging the stance of all responses under the source tweet and removing stance-irrelevant responses, thereby reconstructing the denoised misinformation propagation tree, so that the subsequent model can focus more on key stance clues to achieve a more robust authenticity assessment.

[0101] To achieve the above objectives, this stage introduces an LLM (Legacy Classifier) ​​to classify the stance of all comments associated with the source tweet, dividing each comment into four categories: support, deny, question, and neutral. Considering the varying actual gains of different LLMs in downstream detection tasks, multiple candidate stance classifiers can be selected in practice, and their respective APIs can be called to determine the stance of all responses. Ultimately, the LLM used is determined based on the "impact on downstream detection performance" of the data filtered by these LLMs.

[0102] S2. The adjacency matrix of the denoised propagation tree obtained in S1 is extracted as a structural prior and injected into the attention calculation of the graph Transformer model to model the propagation pattern and obtain the propagation pattern representation; and the semantic features of the source tweet are extracted using a pre-trained language model to obtain the semantic representation of the source tweet.

[0103] In practical implementation, the propagation pattern is modeled to obtain a propagation pattern representation, including:

[0104] For a denoising propagation tree G=(V,E), where the set of nodes is... Including the source tweet and the sequence of user comments ;Will The text content of all nodes in the model is input into the BERT model to obtain the feature embeddings. ;

[0105] The features are embedded into the x-input graph Transformer model, and the self-attention weights of the h-th attention head are... for:

[0106] ;

[0107] in, Indicates the first The self-attention weights of the heads; A is the adjacency matrix of the denoising propagation tree; , The first The query matrix and key matrix of each attention head; Dimensions for each attention head;

[0108] By combining the self-attention output with the residual connection, and then passing it through a feedforward network and pooling operations, a propagation mode representation is obtained. :

[0109] ;

[0110] in, Indicates pooling; Indicates a feedforward network; For the number of heads; This is the value matrix of the h-th attention head; This is the learnable parameter matrix.

[0111] This approach, by directly embedding the graph structure prior (adjacency matrix A) into the attention mechanism's computation process, rather than simply fusing them at the input or output layers, achieves fine-grained coupling of structural and semantic information at the attention level. This allows the model to not only rely on textual semantic similarity when calculating inter-node correlations but also to enforce respect for real-world propagation topologies, such as the temporal order and dependencies between parent and child comments. This effectively alleviates the problems of pure text models ignoring propagation paths and being susceptible to semantic interference. It also outperforms existing methods that only use graph convolutional networks to process structure but lack deep semantic interaction, significantly improving the ability to model dynamic patterns of rumor propagation.

[0112] S3. Input the propagation pattern representation obtained in S2 into the trained fact decoupler to decouple the fact binding features and false position noise features.

[0113] S4. Input the semantic representation of the source tweet obtained in S2 into the trained domain decoupler to decouple the domain-invariant features and domain-specific features.

[0114] In practice, the trained fact decoupling device is obtained through the following training process:

[0115] S3.1 Construct a fact decoupler, the fact decoupler including a first gated branch for extracting fact binding features. and the second gating branch used to extract false position noise features The parameters are respectively ;

[0116] S3.2 Input the propagation pattern representation in the training data into the fact decoupler, and obtain the hidden representation through the self-attention mechanism. ;

[0117] In practical implementation, it is characterized by the propagation model. As input, the hidden representation is first obtained. This method uses self-attention to model the hidden representation:

[0118] ;

[0119] in and It is a learnable matrix.

[0120] S3.3 In the first gating branch, a gating mechanism is used to extract information from the hidden representation. Extract fact binding features from the data, including the gating vector. The formula for calculation is:

[0121] ;

[0122] in, It is the sigmoid function; , , , These are learnable parameters; To correct the activation function of the linear unit; where the superscript d is the hidden layer dimension, and the superscript dp is the projection dimension of the query, key, and value vectors in the self-attention mechanism;

[0123] The gate vector With the hidden representation Element-wise multiplication is performed, followed by layer normalization and multilayer perceptron mapping to obtain fact-binding features. :

[0124] ;

[0125] in, Representation layer normalization; It is a multilayer perceptron;

[0126] S3.4 In the second gating branch, the same gating structure as the first gating branch is adopted, using the gating vector calculated from another set of learnable parameters. Extracting false position noise features ;

[0127] S3.5 Construct the adversarial training objective function for the fact decoupler. :

[0128] ;

[0129] ;

[0130] ;

[0131] in, This indicates that the fact is bound to a feature. Input fraud detector The resulting classification loss; This indicates that the false position noise feature will be used. Input the fake information detector The resulting adversarial loss, where λ is a hyperparameter. , , These are the learnable parameters of the first gating branch, the second gating branch, and the fake information detector, respectively; x∈ y∈ , These represent the sample set and the label set, respectively. The cross-entropy loss function;

[0132] In practice, a gradient inversion layer (GRL) is introduced: forward propagation performs an identity mapping, and backward propagation multiplies the gradient by... This is then passed back to the previous layer, enabling adversarial optimization between the spurious branch and the detector, and allowing for easy insertion between the spurious branch and the detector. The specific formula is as follows:

[0133] ;

[0134] ;

[0135] S3.6 Introduce reconstruction loss constraints to bind the facts to features. With the aforementioned false position noise features After concatenation, the hidden representation is reconstructed using a reconstruction function. Reconstruction loss for:

[0136] ;

[0137] Where [;] represents the vector concatenation operation;

[0138] S3.7, Combine the aforementioned adversarial training objective function and the reconstruction loss Total loss The fact decoupler is trained, and during the training process, an adversarial optimization between the second gating branch and the fake information detector is achieved through a gradient inversion layer until the model converges.

[0139] Through a synergistic mechanism of adversarial decoupling and reconstruction constraints, this scheme achieves explicit separation of factual content and stance noise in the representation of propagation patterns. It ensures that fact-bound features retain core factual clues sufficient to support rumor detection, such as event elements, time, and location, while also eliminating interference signals induced by false stances, such as emotional expressions and misleading rhetoric. This significantly improves the domain robustness and semantic purity of the features. Compared to traditional decoupling methods that rely solely on reconstruction loss or unsupervised clustering, this strategy guides the model to learn more discriminative and thoroughly decoupled factual representations through explicit adversarial objectives, laying a crucial foundation for subsequent cross-domain detection.

[0140] In specific implementation, the trained domain decoupling device is obtained through the following training process:

[0141] S4.1 Construct a domain decoupler, which includes a third branch for extracting domain-invariant features and a fourth branch for extracting domain-specific features;

[0142] S4.2, Semantic representation of source tweets in training data The domain decoupler is input, and using the same gating mechanism, layer normalization, and multilayer perceptron mapping steps as in steps S3.3 to S3.4, domain-invariant features are obtained. Domain-specific features ;

[0143] S4.3 Construct a domain classifier to classify the domain-specific features. Input the domain classifier to predict its domain label and calculate the corresponding classification loss. Introducing reconstruction loss To constrain the domain-invariant features With the domain-specific features Capable of jointly reconstructing the semantic representation of the source tweet back to the input ;

[0144] S4.4, Combining Classification Loss and reconstruction loss Total loss The domain decoupler is trained, and during the training process, an adversarial optimization between the domain-specific feature branch and the domain classifier is achieved through a gradient inversion layer until convergence.

[0145] It is important to emphasize that the goal of a domain classifier is to predict "domain labels," rather than to distinguish between the source and target domains.

[0146] To facilitate a better understanding by those skilled in the art, the following explanation is provided.

[0147] Steps S2-S4 of this method correspond to Figure 2 The Domain Heterogeneity Mitigation Module (DHMM) in the system framework aims to mitigate two types of heterogeneity simultaneously: interdomain heterogeneity caused by differences in propagation semantics between different domains, and clue mismatch heterogeneity caused by the mismatch between propagation patterns and source tweet semantics.

[0148] To mitigate the heterogeneity between the propagation pattern and the semantics of the source tweet, as well as between different domains, DHMM consists of two key components: an encoder layer and a dual-thread domain heterogeneity decoupling layer.

[0149] (1) Coding layer

[0150] The encoder layer consists of BERT and a graph Transformer, used to model the semantics of the source tweet and the propagation pattern, respectively. For a given propagation structure... The sequence of tweets and comments is denoted as First, all text content is input into BERT to obtain the corresponding feature embeddings: To capture stance patterns, following insights from dynamic modeling of rumor propagation, a propagation structure-aware graph Transformer is employed. Let... ( (for hidden layer dimensions), in Head self-attention, Projected to Three spaces: The learnable projection matrix of each attention head satisfies For the first The self-attention weight of the head is calculated using the following formula:

[0151]

[0152] in The adjacency matrix of the propagation structure is used to inject structural priors. Subsequently, the self-attention output is combined with the residual connections, and then processed through a feedforward network and pooling operations to obtain an aggregated representation of the propagation patterns.

[0153]

[0154] in This is the learnable parameter matrix.

[0155] (2) Dual-thread domain heterogeneity decoupling layer

[0156] Domain heterogeneity stems not only from differences in dissemination patterns across different domains, but also from the mismatch between dissemination patterns and the semantic features of source tweets; while existing cross-domain methods... [17,21] Often, focusing only on the former and forcing the learning of domain-invariant representations while neglecting the latter leads to a situation where samples can only rely on unified discriminative features when crossing domains, lacking adaptability. To address this, a dual-clue decoupling layer is proposed, comprising (a) a fact decoupling unit and (b) a domain decoupling unit, to simultaneously alleviate both "inter-domain heterogeneity" and "heterogeneity of propagated semantic mismatch".

[0157] (a) The fact decoupler and (b) the domain decoupler and specific content, as described in steps S3-S4, will not be repeated here.

[0158] S5. After reweighting the fact binding features obtained in S3 and the domain-specific features obtained in S4, dynamic weighted fusion is performed to generate the cue-dominant representation.

[0159] In practice, fact-binding features and domain-specific features are reweighted separately, including:

[0160] The fact binding feature The reweighted mean is calculated using the first multilayer perceptron. and weighted variance And generate fact reweighted features according to the following formula. :

[0161] ;

[0162] in, This indicates element-wise multiplication.

[0163] And using the same calculation method, a domain-specific reweighted algorithm is generated. .

[0164] By introducing learnable feature-level dynamic reweighting, this scheme effectively alleviates the model sensitivity problem caused by excessive differences in the strength of factual features among different samples. For example, weak factual signals such as veiled statements are easily ignored, while strong noise signals such as high sentiment expressions easily dominate the decision. This mechanism adaptively adjusts the contribution of features based on their own distribution statistics, making the model more focused on relatively significant and stable factual components. This further improves the discriminative stability and robustness of features while maintaining the decoupled structure, and is especially suitable for real rumor scenarios with diverse factual expressions and large signal-to-noise ratio fluctuations.

[0165] The process of dynamically weighted fusion to generate clue-dominant representations includes:

[0166] Reweighting the facts With the domain-specific reweighted features As input, the dynamic fusion weights are calculated using the softmax function after mapping through a third-layer perceptron. and :

[0167] ;

[0168] The features are weighted and fused according to the following formula to obtain the cue-dominated representation. :

[0169] ;

[0170] in, This indicates vector concatenation.

[0171] By employing a weighted fusion strategy based on dynamic decision-making of feature content, this approach overcomes the limitations of fixed-ratio fusion or simple splicing. In different samples, the model can autonomously determine the relative importance of fact-binding information and domain-specific context to the current discrimination task, thereby achieving semantic-level adaptive integration of multi-source decoupled features and significantly improving the final representation's ability to focus on the essence of rumors and its adaptability to different scenarios.

[0172] S6. After concatenating the dominant cue representation generated in S5 with the domain-invariant features obtained in S4, input the concatenation into the trained fake information detector for classification and output the rumor detection results. :

[0173] ;

[0174] The fake information detector is a classifier used for binary classification tasks.

[0175] In practical implementation, the overall training objective function of the method is:

[0176] ;

[0177] in, The binary cross-entropy classification loss for the false information detector; The total loss of the fact decoupler; The total loss of the domain decoupler; is a hyperparameter used to balance the detection loss with the training objectives of the two types of decoupled modules.

[0178] To facilitate a better understanding by those skilled in the art, the following explanation is provided.

[0179] Steps S5-S6 of this method correspond to Figure 2The Domain Knowledge Fusion Module (DKFM) in the system framework further addresses the issue that different domains have different degrees of dependence on cross-domain common knowledge and domain-specific knowledge when making decisions. It learns adaptive cue-driven representations, thereby improving the generalization ability under the condition that the target domain is unknown.

[0180] After the dual-clue domain heterogeneity decoupling layer, the model has obtained fact-binding representations and domain-invariant representations. At this point, simply concatenating these two representations and inputting them into the false information detector for prediction, while ignoring domain-specific features, fails to capture the varying degrees of dependence on propagation patterns and domain-specific knowledge in decision-making across different domains. Therefore, the AdaIN method is used to adaptively reweight and fuse domain-specific features and fact-binding features, learning domain-relevant cue-dominant features. This allows for the selective reinforcement of more important knowledge and the suppression of less contributing knowledge across different domains.

[0181] The specific content of the Domain Knowledge Fusion Module (DKFM) is described in steps S5-S6 and will not be repeated here.

[0182] Compared to existing technologies, this invention introduces a large-scale language model to perform fine-grained stance classification of user comments and selectively removes neutral comment nodes, effectively eliminating a large amount of noise interference in the propagation structure. Compared to existing technologies that directly use the original propagation tree, this method not only simplifies the complexity of the graph structure but, more importantly, ensures that subsequent propagation pattern modeling is based solely on stance-related comments with discriminative value, thereby significantly improving the purity and discriminative power of structural features. Furthermore, this solution innovatively injects the adjacency matrix of the denoised propagation tree as a structural prior into the attention calculation of the graph Transformer, while simultaneously utilizing a pre-trained language model to extract deep semantic features of the source tweets. Compared to the simple concatenation or independent processing of structural and semantic features in traditional methods, this fusion strategy enables deep interaction between propagation patterns and textual content, resulting in richer and more comprehensive information representation. Moreover, through the collaborative work of fact decouplers and domain decouplers, this solution achieves a refined decomposition of propagation pattern representation and semantic representation. The fact decoupling decomposes the propagation pattern into fact-bound features and false position noise features, while the domain decoupling decomposes semantic features into domain-invariant features and domain-specific features. This dual decoupling mechanism effectively overcomes the feature entanglement problem in existing technologies, enabling the model to focus on the most essential discriminative information and significantly improving cross-domain generalization ability.

[0183] The dynamically weighted fusion mechanism designed in this scheme can adaptively adjust the fusion weights of fact-binding features and domain-specific features based on the differences in the importance of various features in different domains and specific contexts. Compared with the fixed-weight fusion strategies in existing technologies, this method can more flexibly capture the most discriminative cues, thereby generating more accurate cue-dominant representations and providing strong support for the final rumor detection decision. In addition, by concatenating the cue-dominant representations with domain-invariant features before inputting them into the detector, this scheme achieves an organic combination of domain-specific knowledge and cross-domain general knowledge. This design retains the discriminative ability for specific domains while ensuring the stability of the model when facing unknown domains, effectively solving the problem of insufficient domain adaptability in existing technologies.

[0184] This method can effectively decouple fact-binding features from domain-specific features while removing interference from position-irrelevant comments, thus achieving high-precision cross-domain rumor detection.

[0185] Example 2

[0186] To better illustrate the effectiveness of this method, the following experiment was conducted.

[0187] • Datasets and Settings

[0188] The proposed method was evaluated on two cross-domain rumor datasets. The first dataset, from the Pheme dataset, includes five events, as shown in Table 1. Four of these events were designated as the source domain, and the remaining one as the target domain. The second dataset includes Twitter15, Twitter16, and Twitter19, as shown in Table 2. Twitter15 and 16 were designated as the source domain, and Twitter19 as the target domain.

[0189] Table 1. Detailed data of the Pheme dataset.

[0190]

[0191] Table 2. Detailed data from the Twitter dataset.

[0192]

[0193] Overall performance

[0194] In the comparative experiments, the model's accuracy and the F1 score for each of the two classes of samples were reported on two cross-domain misinformation datasets to comprehensively characterize the model's ability to distinguish between rumor and non-rumor categories. The comparisons covered both SLM-based and LLM-based methods; to ensure the reliability of the results, the experimental results presented in the tables (including each baseline method) are the average of three independent runs.

[0195] First, on the Phome5 dataset, our proposed method (UTBD, which will not be elaborated further) is compared with all baseline methods under five different cross-domain settings, as shown in Tables 3, 4, and 5. Experimental results show that UTBD outperforms the baseline methods in most settings, only slightly lagging behind UCD-RD and MONTROSE in a few. Further analysis reveals that UTBD outperforms the state-of-the-art cross-domain detection methods UCD-RD and MONTROSE in three out of the five settings, while performing slightly worse in two others. It is important to emphasize that UCD-RD and MONTROSE are customized for cross-domain misinformation detection scenarios, typically utilizing unlabeled samples from the target domain for alignment or training to facilitate cross-domain adaptation. In contrast, UTBD belongs to the target domain-agnostic paradigm, relying solely on source domain data during training without accessing any target domain information. Under this more stringent setting, UTBD still achieves overall leading detection performance and obtains the highest average accuracy across all experimental settings, thus validating the effectiveness of UTBD in domain-generalized misinformation detection tasks.

[0196] Table 3 Comparison of cross-domain detection performance in the Sydney Siege (S) and Ottawa Shooting (O) domains.

[0197]

[0198] Table 4. Comparison of cross-domain detection performance on the Germanwings Crash (G) and Ferguson Unrest (F) domains.

[0199]

[0200] Table 5. Cross-domain detection performance on the Charlie Hebdo (C) domain and average performance across five domains.

[0201]

[0202] Secondly, the second group used the Twitter15 and Twitter16 datasets as the source domain and the Twitter19 dataset as the target domain. The experimental results are shown in Table 6.

[0203] Table 6 Comparison of cross-domain detection performance on the Twitter19 dataset.

[0204]

[0205] The results further demonstrate that, under this cross-domain setting, our proposed method also achieved the highest Acc and F1 scores in UTBD. Whether compared to SLM-based or LLM-based fake information detection methods, our method performs optimally, demonstrating its cross-domain robustness under different platform and topic migration conditions, as well as its superior detection performance in low-resource scenarios.

[0206] Furthermore, to analyze the contribution of "position-independent response filtering" to overall performance, the results table also presents the performance of UTBD on the original unfiltered data (denoted as UTBD*). The comparison shows a significant performance gap between UTBD and UTBD*, indicating that a large number of position-independent responses in the propagation tree do indeed interfere with propagation pattern modeling, thus weakening detection performance. Meanwhile, the experimental results of UTBD* show that even without response filtering, UTBD* still achieves better accuracy and F1 scores than most baselines. This further demonstrates that the dual-clue feature decoupling and domain knowledge fusion modules in UTBD provide stable gains for cross-domain detection.

[0207] Finally, the operating efficiency of UTBD with and without the filtering stage was compared, as shown in Table 7.

[0208] Table 7 Comparison of time efficiency of data before and after filtering for each dataset

[0209] Experiments show that introducing position-independent response filtering can reduce the overall model runtime by about 25%, indicating that this stage can not only improve the performance of subsequent detections, but also significantly reduce the computational overhead caused by the redundancy of the propagation structure, thereby improving inference and deployment efficiency while ensuring effectiveness.

[0210] Ablation experiment

[0211] To evaluate the contribution of each component module in the UTBD framework to the overall performance, ablation experiments were conducted, in which key components of the model were removed one by one and the changes in detection performance were observed. Specifically, three ablation configurations were set: (1) w / o FD: removing the fact decoupler; (2) w / o DD: removing the domain decoupler; (3) w / o KF: removing the domain knowledge fusion module. The corresponding ablation results are as follows: Figure 4 As shown.

[0212] The ablation results lead to the following conclusions: First, the model performance consistently declines when any single component is removed, indicating that each module of UTBD has a significant positive impact on cross-domain detection performance. Second, compared to removing the knowledge fusion module, removing the fact decoupler or domain decoupler results in a more significant performance drop, demonstrating that these two types of decoupling modules play a crucial role in UTBD. These phenomena empirically validate two core design principles of UTBD: first, the proposed dual-clue feature decoupling mechanism effectively improves the reliability and transferability of representations in cross-domain scenarios; second, the domain knowledge fusion strategy is equally important for maintaining the robust performance of the model under complex cross-domain conditions.

[0213] • Visualization results

[0214] To verify whether the performance improvement of the UTBD model in cross-domain fake information detection comes from its superior representation learning ability, the T-SNE method is used to visualize the target domain feature representation before the classifier in complex cross-domain scenarios, such as... Figure 5 As shown. And it is compared with the corresponding feature distribution of the baseline model MONSTRSE, as shown. Figure 6 As shown.

[0215] Visualization results show that UTBD significantly outperforms MONSTRSE in terms of feature representation quality and discriminability: the feature embeddings generated by UTBD form more compact and clearly defined cluster structures, with greater separation between different clusters, and can correspond to different ground truth labels; in contrast, the feature cluster distribution of MONSTRSE is more discrete, and there is significant inter-class overlap. These results further validate the effectiveness of the proposed dual-clue feature decoupling mechanism from the perspective of representation space, and emphasize the important role of domain knowledge fusion strategies in achieving robust cross-domain performance.

[0216] • The impact of hyperparameters

[0217] To assess the sensitivity of the UTBD model to hyperparameter settings, the impact of changes in hyperparameters ω1, ω2, and ω3 on the overall model performance was examined on six sets of cross-domain data.

[0218] like Figure 7As shown, these three hyperparameters were adjusted and compared within the interval [0, 1]. Experimental results show that different settings of ω1, ω2, and ω3 have varying degrees of impact on the model's detection performance. When ω1, ω2, and ω3 are all set to 0.2, the detection performance is the lowest. When ω1, ω2, and ω3 are set to 0.6, 0.6, and 1 respectively, UTBD achieves optimal performance under all experimental settings. Furthermore, the parameter ω1 of the main detector part has the greatest impact on the overall performance of the model, with the performance at 0.2 being about 3% lower than that at 0.6; while the parameter ω3 of the domain decoupling part has the least impact on the overall performance of the model, with the performance at 0.2 being about 2% lower than that at 0.6. In addition, it was observed that on the two sets of cross-domain datasets, the hyperparameter performance curves of UTBD show almost identical trends, with relatively small overall fluctuations.

[0219] ·Summarize

[0220] This method (UTBD) first uses LLM to filter position-independent comments in the propagation structure and reconstructs a denoised rumor propagation tree to reduce the interference of response noise on propagation pattern modeling. Then, it uses Graph Transformer to obtain position pattern representations during the propagation process and BERT to extract semantic features of the source tweets. Based on this, a dual-clue domain heterogeneity decoupling layer is further designed: on the one hand, it separates fact-binding features from false position noise in the propagation pattern; on the other hand, it decouples domain-invariant features from domain-specific biases in the semantic representation of the source tweets. Finally, a knowledge reweighting fusion strategy is used to generate cue-dominant representations for different domains, thereby promoting cross-domain knowledge transfer and generalization discrimination. Experimental results based on two sets of publicly available cross-domain datasets show that UTBD achieves optimal detection performance under the strict setting of using only source domain data and keeping the target domain invisible.

[0221] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit the technical solutions. Those skilled in the art should understand that any modifications or equivalent substitutions to the technical solutions of the present invention without departing from the spirit and scope of the present invention should be covered within the scope of the claims of the present invention.

Claims

1. A cross-domain rumor detection method based on domain decoupling, characterized in that, Includes the following steps: S1. Construct an initial propagation tree for the source tweet and its associated user comments; use a large language model to classify the stance of each user comment in the initial propagation tree, with stance categories including neutral and non-neutral; delete comment nodes with a neutral stance category, and for the deleted non-leaf neutral nodes, reconnect all their child nodes to their parent nodes to reconstruct the denoised propagation tree. S2. Extract the adjacency matrix of the denoised propagation tree obtained in S1 as a structural prior and inject it into the attention calculation of the graph Transformer model to model the propagation pattern and obtain the propagation pattern representation; and use the pre-trained language model to extract the semantic features of the source tweet to obtain the semantic representation of the source tweet. S3. Input the propagation pattern representation obtained in S2 into the trained fact decoupler to decouple the fact binding features and false position noise features. S4. Input the semantic representation of the source tweet obtained in S2 into the trained domain decoupler to decouple the domain-invariant features and domain-specific features. S5. After reweighting the fact binding features obtained in S3 and the domain-specific features obtained in S4, dynamic weighted fusion is performed to generate a clue-dominated representation. S6. After concatenating the dominant cue representation generated in S5 with the domain-invariant features obtained in S4, input the concatenation into the trained fake information detector for classification and output the rumor detection results. S3 and S4 can be interchanged.

2. The cross-domain rumor detection method based on domain decoupling according to claim 1, characterized in that, In S1, the non-neutral stance categories include support, opposition, and questioning; The step of classifying the stance of each user comment in the initial propagation tree using a large language model includes: calling the large language model interface, taking the source tweet and a single user comment as input, and using a prompt template to guide the large language model to classify the stance of each user comment into one of the following categories: support, opposition, questioning, or neutral.

3. The cross-domain rumor detection method based on domain decoupling according to claim 1, characterized in that, In S1, the step of reconnecting all child nodes to their parent nodes includes: traversing the initial propagation tree using a breadth-first search strategy; when it is determined that the stance category of the currently traversed comment node is neutral, deleting the neutral node from the initial propagation tree and pointing the parent pointers of all child nodes of the neutral node to the parent node of the neutral node before deletion, until all neutral nodes have been processed, thus obtaining the denoised propagation tree.

4. The cross-domain rumor detection method based on domain decoupling according to claim 1, characterized in that, In S2, the propagation pattern is modeled to obtain a propagation pattern representation, including: For a denoising propagation tree G=(V,E), where the set of nodes is... Including the source tweet and the sequence of user comments ;Will The text content of all nodes in the model is input into the BERT model to obtain the feature embeddings. ; The features are embedded into the x-input graph Transformer model, and the self-attention weights of the h-th attention head are... for: ； in, Indicates the first The self-attention weights of the heads; A is the adjacency matrix of the denoising propagation tree; , The first The query matrix and key matrix of each attention head; Dimensions for each attention head; By combining the self-attention output with the residual connection, and then passing it through a feedforward network and pooling operations, a propagation mode representation is obtained. : ； in, Indicates pooling; Indicates a feedforward network; For the number of heads; This is the value matrix of the h-th attention head; This is the learnable parameter matrix.

5. The cross-domain rumor detection method based on domain decoupling according to claim 1, characterized in that, The trained fact decoupling device is obtained through the following training process: S3.1 Construct a fact decoupler, the fact decoupler including a first gated branch for extracting fact binding features. and the second gating branch used to extract false position noise features The parameters are respectively ; S3.2 Input the propagation pattern representation in the training data into the fact decoupler, and obtain the hidden representation through the self-attention mechanism. ; S3.3 In the first gating branch, a gating mechanism is used to extract information from the hidden representation. Extract fact binding features from the data, including the gating vector. The formula for calculation is: ； in, It is the sigmoid function; , , , These are learnable parameters; To correct the activation function of the linear unit; where the superscript d is the hidden layer dimension, and the superscript dp is the projection dimension of the query, key, and value vectors in the self-attention mechanism; The gate vector With the hidden representation Element-wise multiplication is performed, followed by layer normalization and multilayer perceptron mapping to obtain fact-binding features. : ； in, Representation layer normalization; It is a multilayer perceptron; S3.4 In the second gating branch, the same gating structure as the first gating branch is adopted, using the gating vector calculated from another set of learnable parameters. Extracting false position noise features ; S3.5 Construct the adversarial training objective function for the fact decoupler. : ; ; ; in, This indicates that the fact is bound to a feature. Input fraud detector The resulting classification loss; This indicates that the false position noise feature will be used. Input the fake information detector The resulting adversarial loss, where λ is a hyperparameter. , , These are the learnable parameters of the first gating branch, the second gating branch, and the fake information detector, respectively; x∈ y∈ , These represent the sample set and the label set, respectively. The cross-entropy loss function; S3.6 Introduce reconstruction loss constraints to bind the facts to features. With the aforementioned false position noise features After concatenation, the hidden representation is reconstructed using a reconstruction function. Reconstruction loss for: ； Where [;] represents the vector concatenation operation; S3.7, Combine the aforementioned adversarial training objective function and the reconstruction loss Total loss The fact decoupler is trained, and during the training process, an adversarial optimization between the second gating branch and the fake information detector is achieved through a gradient inversion layer until the model converges.

6. The cross-domain rumor detection method based on domain decoupling according to claim 5, characterized in that, The trained domain decoupler is obtained through the following training process: S4.1 Construct a domain decoupler, which includes a third branch for extracting domain-invariant features and a fourth branch for extracting domain-specific features; S4.2, Semantic representation of source tweets in training data The domain decoupler is input, and using the same gating mechanism, layer normalization, and multilayer perceptron mapping steps as in steps S3.3 to S3.4, domain-invariant features are obtained. Domain-specific features ; S4.3 Construct a domain classifier to classify the domain-specific features. Input the domain classifier to predict its domain label and calculate the corresponding classification loss. Introducing reconstruction loss To constrain the domain-invariant features With the domain-specific features Capable of jointly reconstructing the semantic representation of the source tweet back to the input ; S4.4, Combining Classification Loss and reconstruction loss Total loss The domain decoupler is trained, and during the training process, an adversarial optimization between the domain-specific feature branch and the domain classifier is achieved through a gradient inversion layer until convergence.

7. The cross-domain rumor detection method based on domain decoupling according to claim 1, characterized in that, In S5, fact-binding features and domain-specific features are reweighted separately, including: The fact binding feature The reweighted mean is calculated using the first multilayer perceptron. and weighted variance And generate fact reweighted features according to the following formula. : ； in, This indicates element-wise multiplication. And using the same calculation method, a domain-specific reweighted algorithm is generated. .

8. The cross-domain rumor detection method based on domain decoupling according to claim 7, characterized in that, In S5, the dynamic weighted fusion to generate a cue-dominant representation includes: Reweighting the facts With the domain-specific reweighted features As input, the dynamic fusion weights are calculated using the softmax function after mapping through a third-layer perceptron. and : ； The features are weighted and fused according to the following formula to obtain the cue-dominated representation. : ； in, This indicates vector concatenation.

9. The cross-domain rumor detection method based on domain decoupling according to claim 1, characterized in that, The fake information detector is a classifier used for binary classification tasks.

10. The cross-domain rumor detection method based on domain decoupling according to claim 1, characterized in that, The overall training objective function of the method is: ； in, The binary cross-entropy classification loss for the false information detector; The total loss of the fact decoupler; The total loss of the domain decoupler; is a hyperparameter used to balance the detection loss with the training objectives of the two types of decoupled modules.

Images