A cross-domain fake information detection method

Abstract

A cross-domain fake information detection method, the steps are: first, special word extraction, independent training of domain-text fusion model and domain model, respectively capturing domain-text features and domain features, then subtracting domain features from domain-text features, guiding the model to extract text features independent of specific domain; second, modeling the propagation path graph of tweets, using graph neural networks to capture the propagation path structure features of tweets; then using feature fusion and label prediction module to fuse text sharing features and propagation path graph structure features to get the final comprehensive representation of tweets; finally, by comparing learning to narrow the domain distribution difference between source domain and target domain, aligning the features of tweets in two domains, and using the classifier to complete the detection of target domain fake information. This detection method has greatly improved the accuracy in the task of cross-domain fake information detection, which is of great significance for the security and effective governance of social network space.

Description

Technical Field

[0001] This invention relates to the fields of data mining, natural language processing, and graph neural networks, and more specifically to a cross-domain method for detecting misinformation. Background Technology

[0002] In recent years, with the development of internet technology, online interactive platforms, represented by social media, have rapidly become widespread. While social media provides users with convenient channels for information access and communication, it has also become a breeding ground for the spread of misinformation. The rapid spread of misinformation can trigger social panic, economic losses, and political repercussions, and even threaten public safety. Especially in breaking news events, the speed and scope of misinformation's spread are rapid, necessitating effective detection methods for real-time monitoring and response.

[0003] Early methods for detecting misinformation mainly relied on traditional machine learning and manually designed features. These methods generally included the following steps: (1) screening and extracting features with strong representational power from the training samples; (2) training a classification model on the training dataset based on the screened and extracted features; and (3) using the established model to predict data from outside the training dataset, and determining the authenticity of the data through continuous evaluation and optimization. Researchers used manually designed explicit features (such as message length, user information) or implicit features (such as sentiment features, credibility) to represent misinformation. Although rule-based or template-based detection methods are more intuitive, they are highly dependent on domain knowledge and difficult to extend to cross-domain misinformation detection scenarios.

[0004] With the rise of deep learning technology, neural network models have been gradually introduced into the detection of fake information, which has improved detection efficiency and accuracy. These methods can be divided into detection methods based on tweet content and detection methods based on social context.

[0005] Content-based fake news detection methods primarily construct time-series segments from tweets related to an entire event and then detect fake news by extracting latent semantic features. For example, Chen et al. and Ma et al. divided each event's tweets into fixed-length and variable-length time sequences and input them into recurrent neural networks (RNNs) and their variants to capture time-language features for fake news detection. Furthermore, Yu et al. indicated that RNN-based methods might be biased towards the most recently input tweets, insufficient for early detection of fake news. Therefore, researchers proposed the CNN-based detection model CAMI, which can effectively extract key features from posts and utilize convolutional techniques to achieve more refined feature interactions. Social context-based detection methods mainly embed interactions (such as user comments / likes / shares) or information diffusion structures in social networks into dense vectors for subsequent detection. For example, Liu et al. observed that real news and fake news have different propagation patterns; therefore, they used gated recurrent units (GRUs) and CNNs to extract global and local features from the forwarding sequence for fake news detection. Building on this, Lu et al. introduced graph convolutional neural networks (GCNs) to learn more accurate structural information about news propagation paths. Unlike previous studies, Bian et al. emphasized both the depth and breadth of fake news dissemination, thus proposing the Bi-GCN model to comprehensively describe the spread of fake news on social networks from two perspectives. To systematically integrate features from various aspects and make the model more robust, interpretable, and accurate, Lao et al. proposed a rumor detection model based on linear and nonlinear propagation fields (RDLNP). This model automatically detects rumors propagating in both linear and nonlinear ways using tweet content, social context, and temporal information. RDLNP consists of four parts: Rumor Hybrid Feature Learning (RHFL), Nonlinear Structure Learning (NLSL), Linear Sequence Learning (LSL), and Shared Feature Learning (SFL).

[0006] However, existing methods tend to improve fake news detection metrics on specific datasets and are heavily reliant on the training dataset. Due to variations in event distribution, applying them to unseen domains is impractical. Research on how to leverage knowledge learned from historical events to detect fake news in upcoming events is still limited.

[0007] To address the challenge of cross-domain misinformation detection, researchers have proposed detection methods based on machine learning and adversarial learning. Tolosi et al. first raised the issue of cross-domain misinformation detection in 2016. Their research analyzed fake and genuine tweets in the PHEME dataset and found that most feature distributions were domain-specific; even features expected to be domain-independent exhibited domain-specific distributions, leading to lower model performance in real-world applications compared to experimental results. Later, Castelo et al.'s 2019 research demonstrated the existence of cross-domain shared features and proposed a topic-agnostic classification strategy that identifies misinformation by analyzing network and linguistic features. This method maintains high detection accuracy even when facing the dynamic nature and topic changes of misinformation on social networks.

[0008] With the development of deep learning technology, deep neural network architectures have brought impressive progress to various machine learning tasks and applications, greatly improving the efficiency and accuracy of fake information detection and automatically extracting latent high-level feature representations of tweets. However, most of these methods ignore the problem of domain distribution differences. Experiments by Khoo et al. show that when using the PHEME (Twitter platform dataset) dataset, the detection performance drops significantly when the domains of the training and test sets do not overlap. After re-splitting the dataset, the results return to normal. This also proves from another perspective that existing methods mostly extract domain-specific features, and when faced with unseen domains not present in the training set, the model's generalization ability is poor, failing to achieve the expected detection results. Inspired by cross-domain adversarial network architectures, Wang et al. proposed Event Adversarial Neural Networks (EANN) in 2018. This is an end-to-end framework that can learn domain-invariant features to detect fake information in new domains. The EANN model includes a multimodal feature extractor, a fake information detector, and a domain discriminator. By removing event-specific features through the domain discriminator while retaining shared features, it achieves better generalization ability. However, EANN does not consider the differences in the importance of tweets to the target domain. To address this, Ding et al. proposed the MetaDetector in 2021, which improves detection efficiency by evaluating the impact of source domain tweets on the target domain through a pseudo-domain discriminator. Meanwhile, Yuan et al. also proposed DAGA-NN in 2021, utilizing graph attention neural networks and domain-shared features for cross-domain misinformation detection.

[0009] In practical applications, social media content exhibits domain diversity, and data from new domains often lacks labels. Traditional detection methods perform poorly when faced with unseen domains. Adversarial learning-based methods align the feature distributions of the source and target domains using a domain discriminator, effectively improving the model's generalization ability. However, existing methods largely focus on extracting domain-shared features, neglecting the combined influence of text content and propagation structure. Furthermore, class inconsistencies during feature alignment can lead to instability in the decision boundary.

[0010] Based on the above challenges, this invention combines the ideas of contrastive learning and feature debiasing, integrates text-shared features and propagation path structure features, and effectively mitigates the impact of unlabeled data by semantically aligning source and target domain data, thereby improving the model's detection performance in the target domain.

[0011] The differences compared to existing technologies are as follows:

[0012] Technical Comparison with the Paper "Cross-Domain Rumor Detection Based on Dual-Modal DomainAlignment"

[0013] The paper "Cross-Domain Rumor Detection Based on Dual-Modal Domain Alignment" discloses a cross-domain misinformation detection method based on dual-modality, which includes the following steps: using the TextRank method to identify specific domain keywords in the source and target domain texts, then performing data augmentation based on keyword replacement to obtain modified sentence pairs, and then using a Bi-LSTM network to encode the text features of the sentences; using a Bi-GCN network to obtain the propagation structure features of the tweets, and then using contrastive learning to optimize the structure features; finally, classification is performed.

[0014] This invention proposes a cross-domain misinformation detection method that is both innovative and practical. By extracting and debiasing domain features from the text, it obtains domain-independent general text features, significantly reducing interference from low-quality data and inflexible data augmentation rules, thus enhancing robustness. It captures tweet interaction relationships by constructing bidirectional propagation trees and diffusion trees, and extracts propagation path graph features. After fusing the text features and propagation path graph features, it uses contrastive learning for optimization before classification.

[0015] Unlike the method disclosed in the paper "Cross-Domain Rumor Detection Based on Dual-Modal DomainAlignment", this invention trains a domain-text fusion model and a domain model on the basis of the FLAN-T5-large model for feature extraction and domain bias removal, without relying on the quality of augmented data, thus enhancing the model's generalization ability. After fusing text features and structural features, contrastive learning is used for optimization, which helps the model learn the key features that distinguish false information from real information, thereby improving the accuracy of detection. Summary of the Invention

[0016] To address the above problems, this invention proposes a cross-domain misinformation detection method. It extracts text features based on domain-specific words to obtain domain-neutral text features, extracts information propagation path graph features based on propagation trees and diffusion trees, and then uses contrastive learning to optimize the features after fusing them.

[0017] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0018] A cross-domain method for detecting misinformation includes the following steps:

[0019] (1) A text feature extraction method based on domain-specific word extraction and feature debiasing;

[0020] A text feature extraction method based on domain-specific word extraction is adopted to extract domain-specific text from tweets in both the source and target domains. During training, the original text content is used to train the domain-text fusion model, and the domain-specific text is used to train the domain model. In the inference phase, domain-specific features are subtracted from the original text features to obtain text features free of domain information.

[0021] (2) Propagation structure feature extraction module;

[0022] This module combines the tweet content representation obtained in step (1) and aims to learn the propagation characteristics of tweets from the perspective of tweet propagation and based on the interaction relationship between tweets. During the training process, the propagation tree and diffusion tree of information are constructed from two different directions, top-down and bottom-up. Then, the propagation characteristics and diffusion characteristics of information are extracted from the two directions respectively through the GAT model. Finally, the two characteristics are fused to obtain the complete structural representation of the tweet.

[0023] (3) Feature fusion and label prediction module;

[0024] The shared features of tweet text in the source and target domains obtained in step (1) and the structural features of tweet propagation path graphs in the source and target domains obtained in step (2) are concatenated to obtain a more comprehensive final vector representation of the tweets. The spherical K-means algorithm is used to perform unsupervised clustering of tweets in the target domain to generate predicted labels for tweets in the target domain.

[0025] (4) Cross-domain feature alignment module based on contrastive learning;

[0026] Using the predicted label information and the final vector representation of the tweet obtained in step (3), data of the same category are selected as positive pairs and data of different categories are selected as negative pairs across the source and target domains. Through contrastive learning, the distance between features of different categories is re-edited in the feature space, thereby reducing the difference in domain distribution between the source and target domains and obtaining the final optimized features.

[0027] (5) False information detection module;

[0028] Based on the optimized features obtained in step (4), a multilayer perceptron is used to classify tweets by calculating the authenticity score vector of the tweets, thereby realizing the detection of false information in the target domain;

[0029] (6) System function demonstration;

[0030] Based on the cross-domain misinformation detection model proposed above, and using the PHEME and Misinfdect datasets collected by the platform as a foundation, a prototype model for real-time monitoring of cross-domain misinformation was designed. The prototype system consists of two modules: an online module and an offline module. The offline module is responsible for data preprocessing and model training, while the online module is responsible for providing prediction results based on user input, as well as visualization analysis and display.

[0031] As a further improvement of the present invention, step (1) includes the following specific steps:

[0032] (11) In step (1), the special words extraction of the corresponding sentences in the source domain and the target domain is specifically carried out by using the TF-IDF method without labeled data and pre-training to identify and label domain-related special words in the tweet text of the source domain and the target domain.

[0033] (12) The domain-text fusion model trained using the original text content described in step (1) is trained using the domain text. Specifically, based on the special words of the domain text extracted in (12), a FLAN-T5-large model is trained as the domain model to capture the interference caused by domain information to text classification. Then, based on the original text, another FLAN-T5-large model is trained as the domain-text fusion model. In the inference stage, the original text and the special words corresponding to the original text are paired to form text pairs. By subtracting the special word vector features encoded by the domain model from the original text vector features encoded by the domain-text fusion model, text information based solely on shared features can be obtained.

[0034] As a further improvement of the present invention, step (2) specifically includes the following steps:

[0035] (21) Based on the text content vector representation z of each tweet obtained in step (1) i Assume the propagation tree structure of the tweets is G. i Constructing an information propagation tree in a top-down GAT network Constructing an information diffusion tree in a bottom-up GAT network Use an adjacency matrix to represent the interaction relationships between different tweets;

[0036] (22) LeakyReLU is used as the activation function of the activation layer of the GAT network. Multi-head attention mechanism is used to capture the representation of node i from different angles. Then, the Attention mechanism is used to learn the degree of influence of different tweet nodes on the final result. Finally, the features in the two directions are spliced ​​together to obtain the final structural representation.

[0037] As a further improvement of the present invention, in step (3), vector features of shared features of tweet text and propagation path graph structure in the source domain and target domain are obtained from steps (1) and (2), and then concatenated to obtain fused feature O. i To generate predicted labels for tweets from the target domain, a spherical K-means algorithm is used, with the number of clusters set to the number of classes M. Class prototypes from the source domain are used as initial clusters, and cosine similarity is used to measure target features. The distance between the tweet and the m-th cluster center is used to classify the tweets into two categories.

[0038] As a further improvement of the present invention, step (4) obtains the fusion feature O from step (3). i We select data of the same category from the source domain and the target domain as positive pairs and data of different categories as negative pairs. We use contrastive learning to align the features of tweets from the two domains to obtain the final representation of the tweets.

[0039] Beneficial effects: Compared with the prior art, the present invention adopts the above technical solution and has the following advantages:

[0040] (1) In order to better extract cross-domain features from the text content, we first detect domain-related special words in all tweet text content, extract the domain special words and form text pairs with the original text content, then perform feature extraction through the domain model and the domain-text fusion model, and then perform feature debiasing so that the model can capture shared features that are not related to the domain.

[0041] (2) While considering the extraction of shared features of tweet text content, we also consider the propagation path graph structure features of tweets in the entire propagation process, which captures more shared features of false information in different fields and obtains a more comprehensive final representation of tweets.

[0042] (3) The contrastive learning method is used to align the fused features across domains. By re-editing the distance between features of different categories in the feature space, the difference in domain distribution between the source domain and the target domain is reduced. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the invention;

[0044] Figure 2 This is a model diagram of the propagation path graph structure feature extraction module of the present invention;

[0045] Figure 3 This is a model diagram of the feature fusion module of the present invention;

[0046] Figure 4 This is a model diagram of the label prediction module of the present invention;

[0047] Figure 5 This is a schematic diagram illustrating the generation of cross-domain positive and negative examples of the present invention;

[0048] Figure 6 This is a schematic diagram of cross-domain feature alignment according to the present invention;

[0049] Figure 7 This is a diagram illustrating the overall system framework of the present invention;

[0050] Figure 8 Page image showing the comparative experimental results of the prototype system using the PHEME dataset;

[0051] Figure 9 Page image showing the comparison experiment results of the prototype system using the Misinfdect dataset;

[0052] Figure 10 A real-time monitoring page diagram for the prototype system;

[0053] Figure 11The page diagram illustrates the feature alignment process of the prototype system. Detailed Implementation

[0054] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. The following embodiments are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0055] The schematic diagram of this invention is as follows Figure 1 As shown, the model diagram of the propagation path graph structure feature extraction module is as follows: Figure 2 As shown in the figure, the feature fusion module model diagram is as follows: Figure 3 As shown in the figure, the model diagram of the label prediction module is as follows: Figure 4 As shown in the diagram, the generation of positive and negative examples across domains is illustrated below. Figure 5 As shown in the diagram, the cross-domain feature alignment is illustrated below. Figure 6 As shown.

[0056] like Figure 7 As shown, the cross-domain misinformation detection method of the present invention includes the following steps:

[0057] (1) Text feature extraction method based on domain-specific word extraction and feature debiasing

[0058] The main function of this module is to perform initial representation and shared feature extraction of tweet text content. Before entering the tweet text shared feature extraction module, it is necessary to use the TF-IDF method (which requires no labeled data and pre-training) to identify and label domain-related special words in the source and target domain tweet texts during the text data preprocessing stage.

[0059] Suppose that the text content of a tweet in the domain is Where i represents the i-th tweet text in the source domain, and n represents the number of words in the tweet, with the underlined words being the most significant. This represents domain-specific words in the tweet. After extracting domain features using TF-IDF, the domain features of the tweet will be obtained. Create text pairs by combining the original tweet text with the tweet's domain features. The FLAN-T5-large domain-text fusion model is trained using the original tweet text to capture the text classification under the combined effect of domain information and domain-independent information, including the prediction bias caused by domain information, as shown in Equation (2). The FLAN-T5-large domain model is trained using the tweet domain features to capture the text classification under the effect of domain information only, as shown in Equation (3).

[0060]

[0061] Among them, Odomain For domain characteristics, O domain,general These are the original text features that contain domain characteristics. This is represented as cross-entropy loss.

[0062] By using the hyperparameter α, domain features are subtracted from the original text features to obtain irrelevant features free from domain information interference. Through the representation learning of tweets in this module, the information of each tweet is derived from the initial representation. Learn

[0063] (2) Propagation Structure Feature Extraction Module

[0064] Based on the text content vector representation z of each tweet obtained in step (1) i Assume the propagation tree structure of the tweets is G. i Constructing an information propagation tree in a top-down GAT network Constructing an information diffusion tree in a bottom-up GAT network Use an adjacency matrix to represent the interaction relationships between different tweets:

[0065]

[0066] in, This represents the embedding representation of each node in the l-th layer of the graph neural network, while The specific formula for calculating the attention coefficient is as follows:

[0067]

[0068] in, Let h represent the set of tweet node i and its neighboring nodes. l,i and h l,j Let represent the embedding representations of nodes i and j in the l-th layer graph attention network. LeakyReLU is used to compute the attention between nodes, and it is the activation function used by the activation layer. Each tweet node, after aggregating relevant tweet information, uses the attention mechanism to compute the representation of node i in the (l+1)th layer as follows:

[0069]

[0070] To ensure greater stability of GAT, a multi-head attention mechanism is used, and the final representation of node i is:

[0071]

[0072] In the formula, K represents the number of multi-head attention heads, and || represents the join operation operator. Finally, the Attention mechanism is used to learn the degree of influence of different tweet nodes on the final result, and the calculation formula is similar to formulas (6) and (7). After obtaining the representation of each tweet node in the last layer of the network, the representation H of the propagation tree structure can be obtained through the attention mechanism. TD The calculation process for diffusion features is similar to that for propagation features. Finally, the features from both directions are concatenated to obtain the final structural representation.

[0073]

[0074]

[0075] H = H TD ||H BU (12)

[0076] (3) Feature fusion and label prediction module

[0077] After obtaining the vector representations of the shared features and propagation path graph structure of the tweet text in the source and target domains through steps (1) and (2), the two features can be concatenated to obtain a more comprehensive final vector representation of the tweet. i ,|| is the concatenation operator.

[0078] O i =V i ||H i (13)

[0079] When generating positive and negative examples across domains, it is necessary to input the categories of tweets from the source domain and the target domain. During training, the ground-truth labels for the target domain are unavailable, so this module needs to generate predicted labels for tweets from the target domain. Generating predicted labels for tweets from the target domain requires an unsupervised classification algorithm, and spherical K-means is a widely used K-means clustering algorithm. This module uses cosine similarity between texts instead of Euclidean distance to improve computational efficiency, where the calculation of cosine similarity is shown in formula (14):

[0080]

[0081] Because the spherical K-means clustering algorithm is sensitive to initialization, using randomly generated clusters cannot guarantee the semantic relevance of predefined categories. To address this issue, this method sets the number of clusters to the number of classes M and uses class prototypes from the source domain as initial clusters. Specifically, this method first calculates the centroids of source domain tweets in each category as the corresponding class prototypes, and the initial cluster center C of the m-th class is... m Defined as:

[0082]

[0083] Given features from the target domain, spherical K-means clustering is performed using centers initialized from source domain tweets. Cosine similarity is used to measure target features when determining the label for each target sample. The distance between the m-th cluster center and the target domain. After clustering, the distance between each sample in the target domain and the target domain. Each will have a corresponding prediction label. To ensure the effectiveness of the predicted labels, this method removes low-quality samples from their designated cluster centers and selects samples with higher confidence.

[0084] The algorithm is described in Table 1:

[0085] Table 1 Feature fusion and label prediction algorithms

[0086]

[0087] (4) Cross-domain feature alignment module based on contrastive learning

[0088] This module aims to align features between the source and target domains by forming positive and negative example pairs through contrastive learning, learning shared features across domains. It will then align the i-th sample in the target domain... Features After L2 regularization, it serves as an anchor point, forming a direct pair with samples of the same class in the source domain. Its features are denoted as... This method expresses the cross-domain contrast loss as:

[0089]

[0090] in, For those with the same label in the source domain The number of samples, The sample selection is indicated by sim, where cosine similarity is the calculation function and τ represents temperature. When extracting shared text features, tweets in the source and target domains are projected into the same semantic space. However, when extracting propagation structure features, without the constraint of contrastive loss, the features extracted by the network tend to be domain-specific and cannot be applied to the target domain. For data in the target domain, the contrastive classification loss is also calculated in the same way as in formula (16).

[0091]

[0092] Where, N t This represents the total number of tweets targeting the target domain in a batch. For those with the same label in the source domain The number of source domain tweets. Therefore, this method projects source and target domain tweets belonging to the same category closer to each other than source and target domain tweets belonging to different categories.

[0093] Cross-domain contrastive loss improves performance by using bidirectional feature alignment via tweets from both domains. Finally, the cross-domain contrastive loss is compared with the standard cross-entropy loss L from the source domain. ce Combining these methods, we obtain the final training objective function:

[0094]

[0095] (5) False information detection module

[0096] The main function of this module is to evaluate the authenticity of the original tweets in the target domain based on the final representation of the tweets obtained above. Its main component is a multilayer perceptron, which performs classification and prediction by calculating the authenticity score vector of the tweets.

[0097] (6) System Function Demonstration

[0098] The system's functions are mainly divided into offline and online modules. The online module is primarily used for online misinformation detection, including a results display module, a user interaction module, and a visualization analysis module. The offline module performs model training after data collection and preprocessing.

[0099] The datasets used in this method include the English dataset PHEME and the Chinese dataset Misinfdect, and a MySQL database is used to store the fake information data to be detected.

[0100] The data preprocessing module preprocesses the text and structural information in the dataset. It restores the English abbreviations in the PHEME dataset, converts the text content to lowercase, and filters out stop words using a stop word list. It also filters stop words from the Misinfdect dataset, then segments the text using the jieba word segmentation tool. Finally, it uniformly converts network elements in the text content into a consistent representation, such as converting links to... <url>@user converted <user>Words starting with # are uniformly converted to <hashtag>Emojis converted to <emoji>To handle the propagation structure, a top-down propagation tree and a bottom-up diffusion tree are constructed, with the original tweet as the root node, and the propagation structure is processed based on the relationship between retweets and comments.

[0101] In the model training module, for each detection task in the PHEME dataset, this method selects 800 source domain tweets with true / false labels and 800 target domain tweets without true / false labels for training, 100 target tweets for validation, and 200 target tweets for testing. For each detection task in the Misinfdect dataset, this method randomly selects 1300 source domain tweets with true / false labels and 1300 target domain tweets without true / false labels for training, 300 target domain tweets for validation, and 400 target domain tweets for testing.

[0102] The online module primarily provides users with real-time detection capabilities and performs real-time visualization analysis of user input. Through this module, users can obtain experimental results on two benchmark datasets, thereby gaining a deeper understanding of the detection capabilities of the model proposed in this invention. The results are shown below. Figure 8 , Figure 9 More importantly, users can input the relevant Weibo IDs to be detected through the system's provided interface. After receiving the tweet IDs, the system will crawl the specific tweets from the corresponding websites and return the model's detection results. In addition, the system will also perform keyword detection based on the specific content of the tweets and collect some tweets from related fields. Based on the collected tweets, the system will perform tweet length analysis, sentiment polarity analysis, and high-frequency word cloud analysis. The results are shown below. Figure 10 To allow users to more intuitively experience the process of cross-domain misinformation detection, the system will also dynamically display the feature transfer process, as shown in the following demonstration: Figure 11 .

[0103] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any modifications or equivalent changes made based on the technical essence of the present invention shall still fall within the scope of protection claimed by the present invention.< / emoji> < / hashtag> < / user> < / url>

Claims

1. A cross-domain method for detecting misinformation, characterized in that: Includes the following steps: (1) A text feature extraction method based on domain-specific word extraction and feature debiasing; A text feature extraction method based on domain-specific word extraction is adopted to extract domain-shared features from tweet text. The domain-shared features are used to train a domain model, and the original text is used to train a domain-text fusion model. During the inference stage, the domain features are subtracted from the domain-text features to obtain a text feature vector without domain information. (2) Propagation structure feature extraction module; This module combines the tweet content representation obtained in step (1) and aims to learn the propagation path graph features of tweets from the perspective of tweet propagation and based on the interaction relationship between tweets. During the training process, the propagation tree and diffusion tree of information are constructed from two different directions, top-down and bottom-up. Then, the propagation features and diffusion features of information are extracted from the two directions respectively through the GAT model. Finally, the two features are fused to obtain the complete structural representation of the tweet. (3) Feature fusion and label prediction module; The shared features of tweet text in the source and target domains obtained in step (1) and the tweet propagation structure features in the source and target domains obtained in step (2) are concatenated to obtain a more comprehensive final vector representation of the tweets. The spherical K-means algorithm is used to perform unsupervised clustering of the tweets in the target domain to generate predicted labels for the tweets in the target domain. (4) Cross-domain feature alignment module based on contrastive learning; Using the predicted label information and the final vector representation of the tweet obtained in step (3), data of the same category are selected as positive pairs and data of different categories are selected as negative pairs across the source and target domains. Through contrastive learning, the distance between different categories is re-edited in the feature space, thereby reducing the difference in domain distribution between the source and target domains and obtaining the final optimized features. (5) False information detection module; Based on the optimized features obtained in step (4), a multilayer perceptron is used to classify tweets by calculating the authenticity score vector of the tweets, thereby realizing the detection of false information in the target domain; (6) System function demonstration; Based on the cross-domain misinformation detection model based on contrastive learning proposed above, and using the PHEME dataset and Misinfdect dataset collected from the platform as a basis, a prototype model for real-time monitoring of cross-domain misinformation is designed. The prototype system consists of two modules: an online module and an offline module. The offline module is responsible for data preprocessing and model training, while the online module is responsible for providing prediction results based on user input and displaying visualization analysis.

2. The cross-domain false information detection method according to claim 1, characterized in that: Step (1) includes the following specific steps: (11) In step (1), special words are extracted from the corresponding sentences in the source domain and the target domain. Specifically, the TF-IDF method without labeled data and pre-training is used to identify and label domain-related special words in the tweet text of the source domain and the target domain. (12) Step (1) uses the domain shared features to train the domain model and the original text to train the domain-text fusion model. Specifically, two T5-large models are trained respectively to perform text classification tasks on the content containing only domain text information and original text information.

3. The cross-domain misinformation detection method according to claim 1, characterized in that: Step (2) specifically includes the following steps: (21) Based on the text content vector representation z of each tweet obtained in step (1) i Assume the propagation tree structure of the tweets is G. i Constructing a propagation tree of information in a top-down GAT network Constructing an information diffusion tree in a bottom-up GAT network Use an adjacency matrix to represent the interaction relationships between different tweets; (22) LeakyReLU is used as the activation function of the activation layer of the GAT network. Multi-head attention mechanism is used to capture the representation of node i from different angles. Then, the Attention mechanism is used to learn the degree of influence of different tweet nodes on the final result. Finally, the features in the two directions are spliced ​​together to obtain the final structural representation.

4. The cross-domain false information detection method according to claim 1, characterized in that: In step (3), vector features of shared features and propagation structure of tweet text in the source and target domains are obtained from steps (1) and (2), and then concatenated to obtain the fused feature O. i To generate predicted labels for tweets from the target domain, a spherical K-means algorithm is used, with the number of clusters set to the number of classes M. Class prototypes from the source domain are used as initial clusters, and cosine similarity is used to measure target features. The distance between the tweet and the m-th cluster center is used to classify the tweets into two categories.

5. The cross-domain misinformation detection method according to claim 4, characterized in that: Step (4) obtains the fusion feature O from step (3). i We select data of the same category from the source domain and the target domain as positive pairs and data of different categories as negative pairs. We use contrastive learning to align the features of tweets from the two domains to obtain the final representation of the tweets.

Images